agentic-recall

Long-term memory for agentic tasks. It tells you when it doesn't know.

An MCP server that gives Claude Desktop and Claude Code a persistent memory across sessions — and, unlike a plain vector store, an explicit answer when the corpus does not contain what you asked for.

// ~/.claude.json  (Claude Code)  or  claude_desktop_config.json  (Claude Desktop)
"mcpServers": {
  "memory": {
    "command": "npx",
    "args": ["-y", "agentic-recall@2"],
    "env": { "MEMORY_DIR": "/absolute/path/to/your/memory/folder" }
  }
}

That is the whole install. MEMORY_DIR is the one path you supply, and that is deliberate: the server does not go looking through your disk for your notes.

🟥 Note the @2 — pin the major version. npx -y agentic-recall with no version asks npm for whatever is tagged latest every time its cache is cold, so a new major arrives silently and a config that worked yesterday breaks today. That is not hypothetical: 2.0.0 moved five actions to a second tool, and a tester's in-flight session started failing with received 'index' at action mid-task — because npx had re-resolved latest underneath a running workflow. The cache is what makes an unpinned spec feel stable, and a cache is not a pin: it is cleared by npm cache verify, by disk cleanup, and by every new machine.

@2 takes fixes and features, never a breaking major. agentic-recall@2.0.0 pins exactly.

🟥 A warm npx cache does NOT pick up a fix on its own. Measured 2026-09-22: 2.0.3 had been on the registry for over an hour, and three servers spawned with npx -y agentic-recall@2 still ran the cached 2.0.1. To take a fix, clear the cached copy and re-resolve once — macOS/Linux: rm -rf ~/.npm/_npx && npx -y agentic-recall@2 --version; Windows: the Remove-Item line below — then restart your client. serverVersion in every response tells you which version actually answered.

🟥 Windows users: use this config instead. Bare npx does not work for a Node MCP client on Windows. This is measured, not inferred — CI spawns the published package on windows-latest every push, and reports bare npx -> FAILED: spawn npx ENOENT while cmd /c npx connects:

json

Copy

"memory": {
  "command": "cmd",
  "args": ["/c", "npx", "-y", "agentic-recall@2"],
  "env": { "MEMORY_DIR": "C:/Users/<you>/Documents/memories" }
}

On Windows npx is a batch shim (npx.cmd), and Node's spawn will not resolve it without shell: true. A Node-based MCP client using the config above gets spawn EINVAL and the server silently never connects — it appears configured and does nothing. This is the standard Windows MCP gotcha, not specific to this server (claude-code#58510). npm i -g is not an escape — that installs a .cmd shim too. The only shim-free form is pointing node at the file directly: "command": "node", "args": ["C:/Users/<you>/AppData/Roaming/npm/node_modules/agentic-recall/index.js"].

Whether Claude Desktop itself is affected is still unconfirmed — it may spawn through a shell. The cmd /c form works either way, so it is the one to use on Windows.

🟥 Windows: if npx fails with ECOMPROMISED / Lock compromised, clear the npx cache. Node 24/25 with npm 11 leaves a stale lock in npx's cache directory that it then refuses to refresh (npm/cli#8710). Nothing to do with this server — it hits MCP Inspector and Gemini CLI identically. Measured on a real Windows machine: it reproduced twice, and clearing the cache fixed it (then a 98 s first download):

powershell

Copy

Remove-Item -Recurse -Force "$env:LOCALAPPDATA\npm-cache\_npx" -ErrorAction SilentlyContinue
npm cache verify
npx -y agentic-recall@2 --version

If it still fails, two alternatives that avoid npx entirely — both verified end to end: npm install -g agentic-recall (then "command": "agentic-recall", "args": []), or Node 22, where npx works normally.

Requires Node 20 or newer. Nothing else — the embedding model downloads on first index and then runs locally. Your memories never leave the machine.

Prefer to read the code first, or run the suite?

git clone https://github.com/dfrancislyondflabc-tech/agentic-recall.git
cd agentic-recall && npm install && npm test    # 268 checks, no corpus of your own needed

Built by Daniel Francis-Lyon — questions, bug reports and criticism all welcome, either as an issue on this repo or at danfrancislyon@gmail.com.

Privacy Policy

Nothing leaves your machine. Memories are plain files in a folder you choose; the embedding model runs locally after a one-time download; there is no telemetry, no analytics and no phone-home. The author receives no data from your installation and has no means of doing so.

Full text: PRIVACY.md.

The part most memory tools skip

Retrieval always returns something. Ranked by similarity, the least-bad match comes back looking like an answer, and an agent acting on it cannot tell the difference between "here is what you wrote" and "here is the closest thing to it".

This server separates those two cases and says which one you got:

// asked about something that isn't in the corpus
{ "results": [],
  "confidence": "low",
  "bestWeak": [{ "name": "wheel-truing", "score": 0.31 }],
  "absenceNote": "no strong match — 'afternoon' appears nowhere in this corpus" }

Nothing was invented, nothing was silently ranked into first place, and the nearest thing is offered as a candidate rather than as an answer. Everything below is about how that verdict is reached and how it is measured.

The same refusal, applied to engineering claims

A memory saying "I fixed that in abc123" is a sentence. Classifying sentences has failed here three times — a correction vocabulary that fired on 76% of exchanges, an unresolved-statement vocabulary that fired on 24%. But a SHA is not a sentence to classify. It either exists or it does not.

Point MEMORY_GIT_REPOS at your repositories and every hex-shaped token in a retrieved memory is looked up. Four commands, and only what they prove:

Measured over 2,319 ingested exchanges: 707 SHA-shaped candidates collapsed to 355 real commits. Half of what looks like a commit isn't one, which is why every token is checked rather than trusted for its shape. Those rejects are not invented commits: re-measured on the current, larger corpus, 79% of reject occurrences are hex prefixes of conversation session IDs that the deliberately loose 7-10 character pattern sweeps up, so this gap measures how little the shape tells you rather than how often a memory fabricates a SHA.

What this does not tell you, deliberately. It does not check whether the change was later reverted, or whether the thing still exists at HEAD. A commit that landed and was undone the next day still reports as landed. So this is not a claim that the memory is currently true — it is a verified fact about the past, plus rev-list telling you exactly how much has happened since that nothing in the corpus can know about. The gap is measured rather than closed, which is the same discipline as the absence verdict above: never imply knowledge you do not have.

Repos are configured, never inferred — an earlier draft counted commits "in this repo", meaning wherever the process happened to be running, which is a different repository from the one the memories are about.

Two-tier hybrid retrieval over Claude's persistent memory corpus, exposed to Claude Desktop and Claude Code as two MCP tools: memory, which is read-only and never writes to your memory folder, and memory_write, which is the only way anything this server does can change it.

The curated corpus is the folder you point MEMORY_DIR at (see Environment overrides). MEMORY.md, if you have one, is treated as a hand-curated tier-1 index.

Figures quoted throughout this README — "121 .md files, ~2.7 MB", timings, hit rates — are measurements of the author's own corpus, not properties of yours. They are here because a claim with a number behind it can be checked; treat them as the conditions a result was obtained under, not as promises about your data. Three more work corpora sit beside it, each with its own index and its own statistics: other projects' memory folders (projects), auto-ingested conversation exchanges (staging) and the institutional handoff documents (handoff, read-only) — plus the library: imported reference material (books, manuals, policies) in per-category corpora that are searched only when named (scope:'books', or scope:'everything') and can never touch a work score. See Four work corpora + the library, one index each.

Why it exists

Reading the whole memory folder into context costs ~2.7 MB and buries the relevant three lines. This server answers the question "which memories matter for what I'm doing right now?" with three retrievers whose failure modes cancel:

Scores are normalised, fused 42/42/16, then adjusted by a long-document correction, the hot-tier boost and a mild recency decay. Every result reports its provenance (keyword / semantic / phrase / both) so a surprising rank is diagnosable rather than mysterious — and when nothing matched well enough, search says so instead of guessing (see below).

v1.1 — the three defects the 2026-08-14 benchmark measured

retrieval-features-baselines/memory-systems-benchmark.md scored this server against a pre-loaded MEMORY.md over 32 probes and found three real defects. All three are fixed; the numbers are in that file's v1.1 section.

1. BM25 was blind to bodies. It indexed title + description + headings only, so a distinctive phrase living in a body paragraph had to be recovered by the dense leg — structurally the wrong tool for a literal string. The body is now a fourth BM25F field (its own length normalisation, weight 0.3), and a phrase-proximity leg reads the body directly. Verbatim recall went from 4/6 found and 1/6 sentence-located to 6/6 and 6/6, every one at rank 1. Snippets are now cut around the matching window rather than the top of the document, so a quote search returns the sentence you quoted.

2. It could not say "nothing". search always returned limit results; on the four absent probes it handed back confident-looking wrong documents, one at 0.75 — higher than 20 of the 28 correct answers. It now returns noStrongMatch: true with the candidates moved to bestWeak, on 4/4 absent probes with 0/28 false absences. There is no clean score threshold — the distributions overlap across their whole middle — so the verdict is a conjunction of measured weaknesses plus a vocabulary test. the derivation and its margins are recorded in the author's test/ tree.

   🟥 The absence verdict gets less reliable as your corpus gets smaller, and a new corpus is small. orphanShare asks what fraction of your question's distinctive words appear nowhere in the corpus — so on a thin vocabulary, ordinary synonyms are genuinely absent and a question the corpus CAN answer gets refused. Measured here on 122 files: 5 of 20 answerable questions called absent. Measured by an independent reviewer on 13 files: 3 of 4. It fails safe — the right document is in bestWeak, not invented — but on a young corpus read bestWeak before believing a refusal, and expect this to improve as you write more. test/… paths in this README are citations to where a number was measured, not files in this distribution — the suite is not shipped, because it asserts against one private corpus. See CONTRIBUTING.

3. One enormous document was winning everything. A 616 KB changelog took a top-3 slot on 21 of 32 test questions — questions about deployment, about pricing, about a bug in a scraper. It had no business in most of them.

   The cause is in how a document is scored. Documents are split into chunks, and a document's score is the score of its best chunk. That changelog splits into 517 chunks; the typical memory in the corpus splits into 4. So the long document gets 517 chances to have one paragraph that happens to sit near your question, and the short one gets 4. Ask about anything and something in 616 KB of release notes is vaguely on topic.

   It's the same effect as a library where one book runs to 3,000 pages and everything else is a five-page note. Ask any question and the huge book contains a page that looks relevant — not because it is the best answer, but because it had the most chances to match.

   The giveaway was keywordScore: 0 on almost every one of those 21 hits: none of the words in the question appeared in the document at all. It was winning purely on one chunk out of 517 landing near the question in embedding space.

   The fix is to shrink the semantic score of documents that are far longer than that corpus's own normal length. "Normal" is measured per corpus rather than hard-coded, since a corpus of books and a corpus of notes disagree about what long means. And the shrink is waived in proportion to keyword evidence: if your words really are in the document, the penalty lifts. Long is only suspicious when the document didn't match what you actually asked.

   Result: top-3 appearances 21 → 0. It is de-prioritised, not hidden — ask for that changelog by name and it still comes back at rank 1 with keywordScore: 1.0, because now the words match.

The keyword score is on an absolute scale

Fusing two retrievers only works if both scores mean the same thing on every query. A per-query-max normalisation does not: it hands 1.0 to whatever scored best, so on a paraphrase where nothing really matched, an accidental match still carried half the fused score. (A question about which zip packages to maintain returned an unrelated note about a download link at #1, on the token download alone.)

So the keyword leg is scored against measured reference points instead — a raw noise floor, the score a genuine lexical match earns, and the share of the query the document actually answered — each capped by what the query can possibly achieve, so a short exact query like MEMORY is not punished for having little to match. npm run measure-keyword-scale re-runs the measurement those constants came from (four query populations: title-literal, description-literal, in-domain paraphrase, out-of-domain) and prints where the shipping constants sit against it. The derivation lives next to the numbers in lib/config.js.

The fused path only. In bm25-only mode there is no second score to be comparable with, so the per-query-max form is kept and degraded-mode ranking is unchanged.

Install

Did you get this as a zip? Then none of this section applies: open START-HERE.txt and double-click SETUP-WINDOWS.cmd or SETUP-MACOS.command. A zip carries its dependencies and its search model already — there is nothing to clone and nothing to npm install. (Up to 1.7.1 the zips also shipped a dist/ install guide that said the opposite; it is gone.)

From npm — the short way. Nothing to clone and nothing to build; paste the config block at the top of this README into your Claude config and you are done.

npx -y agentic-recall@2 --version     # optional: fetch it now rather than on first launch

From source, if you would rather read it first or run the suite:

git clone https://github.com/dfrancislyondflabc-tech/agentic-recall.git
cd agentic-recall
npm install

Node 20 or newer. There is nothing to build.

Where it keeps its 35 MB. A clone keeps the model cache, the vector cache and the index beside the code, as it always has. A package install (npx, npm i -g) writes them to ~/.agentic-recall instead, because npm's npx cache is disposable and re-downloading the model on eviction would be miserable. MEMORY_ROOT overrides both. Your memories themselves are never in either place — they stay wherever MEMORY_DIR points.

Not the recall-mcp on npm. That name belongs to a different project (a different self-hosted memory server). This one is published as agentic-recall.

Then tell it where your memories are. It does not search your disk for them — there is no sensible default, so it does not guess. MEMORY_DIR (or memoryDir in local-config.json, copied from local-config.example.json; gitignored, never indexed) is the whole of the discovery logic.

If you already use Claude Code's memory, you are done in one line. Point it at that folder and it indexes those files in place — nothing is copied, nothing is converted, and Claude carries on writing them as it always did:

export MEMORY_DIR=~/.claude/projects/<project-slug>/memory

(The slug is your project's path with the separators replaced — ls ~/.claude/projects to find yours.) Reading Claude's own directory, rather than a copy of it, is the reason this server's corpus cannot silently drift out of date.

If your history lives somewhere else, import it once. A ChatGPT export .zip, a folder of Obsidian/Notion markdown, or a single file:

node scripts/import-memories.js /absolute/path/to/export.zip --dry   # preview
node scripts/import-memories.js /absolute/path/to/export.zip

This one does write files into MEMORY_DIR, converting as it goes. memory_write({action: "import"}) is the same thing from inside a conversation.

Remembering the conversations themselves is controlled by the connector toggle, and there is nothing else to configure. While the connector is on, the server spawns a capture walk every five minutes by itself. The only thing a hook adds is the FINAL exchange of a chat you walk away from, captured as it ends rather than on a later sweep — node packaging/install-hooks.mjs adds scripts/auto-ingest.js on Stop and SessionEnd for this install (it backs up ~/.claude/settings.json first, touches no other hook, does nothing on a second run, and reverses with --uninstall). After that the switch you already use in Claude's UI is the switch: connector on means this server is running, which it advertises by leaving a dated mark on disk; the hook reads that mark and captures the session. Connector off, no mark, nothing captured, silently. Captured conversations land in a separate staging corpus at a lower tier, so they are searchable but never outrank a memory you wrote deliberately.

Two overrides live in local-config.json — and it has to be that file rather than an environment variable, because hooks are spawned without your shell environment:

{
  "memoryDir": "/absolute/path/to/your/memories",
  "captureAlways": true
}

captureAlways: true remembers every session, connector on or off. autoIngest: false remembers none, ever. Set one or neither — the default, with both absent, is "remember the sessions you had the connector on for".

And if you had it switched off and only realised afterwards that the work mattered, nothing is lost — the transcript was on disk the whole time, it simply was not ingested. Ask for it after the fact:

memory_write({action: "capture", sinceMinutes: 60})   // remember the last hour
memory_write({action: "capture"})                     // remember this whole session

Already-captured exchanges are skipped, so running it twice is safe. It answers with two counts, because on a live chat the common outcome is neither "new" nor "nothing": exchangesCaptured is exchanges that were not in the store at all, and exchangesRefreshed is exchanges already there whose text had GROWN — normally the turn you are in the middle of. sinceMinutes is measured against each exchange's last activity, so a turn that has been running for an hour is still inside sinceMinutes: 30.

A memory is just a markdown file with a name: and a description: in its frontmatter:

---
name: freehub-service-log
description: Symptoms and fix for the loaner-wheel freehub pawls disengaging under load
---

The pawls stop engaging when the grease thickens, usually on a climb — the cranks turn and
the wheel does not. Strip and re-grease with a light oil, not the heavy grease in the tub.

Related: [[wheel-build-notes]]

name is how the memory is addressed (get, [[wikilinks]]); description is what a search sees first, so it is worth writing as the sentence you would want back. Everything after the frontmatter is the body. Nothing else is required — no metadata: block, no tier, no id. Then:

npm run index        # first build downloads the embedding model, then embeds: ~3 min
npm test             # self-contained: builds its own fixture corpus, needs none of yours
npm run verify       # the same check under its other name

npm test drives the real server over raw stdio JSON-RPC and exercises all thirteen actions against a temporary corpus it writes itself, so it is meaningful on a machine with no memories at all. The exit code is the verdict. (The author's full suite is not public — it asserts against one particular corpus and would fail for you. See CONTRIBUTING.md.)

Measurement scripts, read-only and re-runnable — the tuning constants in lib/config.js each cite the measurement they came from:

npm run measure-keyword-scale   # the absolute keyword scale
npm run analyse-queries         # what has been asked of it

The tool

One gateway tool, thirteen actions: search, get, neighbors, latest, thread, verify, index, index_status, probe_status, promote, demote, import, capture. The four you will use daily are documented in full below.

memory({action: "search", query, limit?, scope?})

Hybrid retrieval. Returns name, description, tier, score, keywordScore, semanticScore, phraseScore, provenance, snippet, links, path, readOnly, corpus (which of the four indexes answered), the attribution block (account, project, sessionId, sessionTitle, type), plus a top-level mode (hybrid | bm25-only | unavailable) and, when degraded, a degradedReason. Default limit 8.

Every response says when its index was built — see Freshness below:

A result's modified is the file's mtime AT INDEX TIME, not a live read. memory({action: "get"}) returns a live stat, as liveModified. Reading one as the other is what produced a confidently wrong conclusion about project state on 2026-08-19.

Search always covers both tiers. Archived memories are fully searchable; they simply do not get the hot boost.

A single document takes at most one top-N slot (RETRIEVAL.maxSlotsPerDoc).

When nothing matched, it says so. Instead of ranking the least-bad document, search can return:

{
  "noStrongMatch": true,
  "confidence": "low",
  "signals": { "topScore": 0.40, "topPhrase": 0.25,
               "lexicalCoverage": 0.24, "orphanShare": 0.46 },
  "absenceNote": "No strong match: the term(s) that make this question specific
                  appear NOWHERE in the corpus (kubernete — 46% of the query's
                  discriminative weight, floor 40%). …",
  "results": [],
  "bestWeak": [ /* the nearest documents, NOT answers */ ]
}

results is emptied and the candidates move to bestWeak, so a caller that checks results cannot accidentally report a non-answer, while one that wants to overrule the verdict still has the candidates. signals is always present on the fused path, verdict or not, so a surprising call is diagnosable.

Two independent rules produce it (constants and margins: measured, not guessed):

• vocabulary — the terms that make the question specific exist nowhere in the corpus (orphanShare ≥ 0.40) and nothing holds the remaining words together. Compound forms are forgiven: de-duplication is absent while duplication is present, and the corpus clearly knows the concept.
• evidence — every word is familiar, but nothing scored (< 0.38), nothing is phrased that way (< 0.40), and most of the question went unanswered (lexical coverage < 0.20). All three, because each alone fires on real questions.

The verdict is advisory and calibrated on this corpus. Never use it to prove a negative that matters — on a held-out set of deliberately vocabulary-free in-domain questions, 5 of 20 are called absent. grep proves a negative; a score does not.

Never claimed in bm25-only mode: the constants are calibrated on the fused three-leg score, so degraded mode reports confidence: "unrated" instead.

What a multi-corpus response is made of

scope: "all" (and an array scope) returns one ranked section per corpus under groups, because the corpora do not share statistics and blending them measurably costs recall. Three rules govern the shape of that envelope; all three came out of one real response, measured on 2026-09-07, which was 56,550 bytes for ten results and told its reader the opposite of the truth.

brief: true on search or latest trims the rows — name, corpus, score, snippet, provenance and the timestamp, nothing else. The envelope is untouched: indexStale, staleWarning, recencyVoid, recentUnindexed, uncapturedSessions, captureHealth, configWarning and guidance are all still there. A caller asking for fewer bytes has not asked to be told less about what was never read.

Measured on the reporting caller's own query against the same corpora, read-only, before and after: 56,634 → 26,842 bytes, and 17,123 with brief: true. The ten ranked rows are 12,426 bytes of that, in both runs — they are the answer, and they are what is left.

Query expansion — queries?, expand? (on by default since 2.1.0; can be turned off completely)

A three-position switch: on (default) | shadow | off. latest ignores it.

To turn it off completely, any one of these (the first that is set wins):

• per call: expand: "off" (or false)
• for the server: MEMORY_QUERY_EXPANSION=off in the connector's env
• for the machine: "queryExpansion": "off" in local-config.json

off adds nothing to any response and writes nothing to the log. An unknown value falls back to the default.

How it is meant to be used: search first, rephrase only on a miss. The server never invents synonyms and never re-queries by itself; the caller owns the guessing. A first search is sent with query alone and costs exactly what it did before. If that search is refused, the response invites a second try. If it answered but the rows do not actually answer the question, the caller judges that and tries again the same way:

• queries: [...] — up to 4 alternative phrasings of the same question, ranked in one call beside query. Each is judged on its own words; a phrasing that refuses adds nothing. results stay the question as asked — same rows, order and scores — each gaining matchedVia, the phrasing(s) that also reached it. Rows only a phrasing reached come back under viaVariants.results with queryScore (what the original question gave them) and alsoBestWeak when the row was already the nearest weak neighbour.
• The verdict stays on query. If the question as asked has no strong match, noStrongMatch stays true and results stays empty, whatever the phrasings found. Read a viaVariants row before relying on it, and say it came from a rephrasing. On a bm25-only index expansion does nothing and says why.

Why the second try is left to the caller: the server cannot see a miss. In the evaluation below, 15 of 17 wrong answers came back at confidence: high, so a server-side "retry when unsure" rule would have kept only 2 of the 7 recoveries.

Measured on a real 4,700-document corpus (pre-registered; 90 answerable questions written by agents that never searched, 20 absent controls): with 3 phrasings, 7 of the 20 questions the first search missed were recovered into viaVariants, 0 results changed, 0 controls answered because of it. Cost per call with phrasings: ~+50 ms on 500 documents, ~+370 ms on 4,200, ~+1.6K tokens. A first search with no phrasings costs nothing extra.

requeryHint is off by default. On an empty result it lists words the nearest documents use and the query did not. It pointed at the right memory 0 times in 3 real refusals, because on a true absence the nearest documents are the wrong ones. It is still computed and logged; turn it on with MEMORY_REQUERY_HINT=on or "requeryHint": "on" in local-config.json.

shadow computes all of this, records it on the query-log row (expansion, redacted like q), and returns the unexpanded baseline plus an expansion: {mode: "shadow"} marker.

memory({action: "latest", query, limit?, scope?, sessionId?, account?, project?})

For state questions — "did X finish", "what happened after Y", "where did we leave X".

search ranks by relevance, and relevance cannot separate "we are starting X" from "X is finished": both are equally about X. That is not a ranker that needs improving, it is the wrong axis for the question. latest filters on every term (no ranking at all) and orders newest first.

It exists because of a specific failure. A session was asked whether a re-parse had completed. It searched, got the exchange where the work started at score 0.88, saw no completion ranked above it, and reported the answer unknowable. The answer was in the corpus the whole time, one term-filter away.

memory({action: "latest", query: "reparse"})
  -> orderedBy: "ts", totalMentions: 16, scopeHint: {scope: "staging", explicit: false}
     results[0]: x-fb357616-20260903T235959000Z   threadPosition: "650 of 650"   laterInThread: 0

Fields worth reading

Query it with identifiers, not prose. This is a literal string filter. Measured over six real questions with known answers:

Same corpus, which held every answer the whole time. Commit SHAs, file names, flags, function names, error strings and exact numbers work; prose belongs in search, which ranks. When the strict filter finds nothing it relaxes to the best available match and sets relaxed + droppedTerms — a dropped term is often the one that mattered.

Compaction summaries are demoted, not dropped. When a session runs out of context the harness reopens it with a summary of everything so far, and that gets ingested as an exchange (34 of 2,318 here). It restates a whole conversation, so it matches almost any filter while carrying a recent timestamp for old content — one took first place on 5 of 6 test questions, once purely because it restated the question. Excluding them outright measured worse: it fixed one question and broke another whose answer existed only inside a summary. So they are labelled isCompactionSummary, sorted below first-hand exchanges, and removable with includeSummaries: false.

scope: "all" returns one section per corpus, never a merged list. The corpora do not share a clock — staging documents carry ts, curated documents carry none — and merging them by time compares incomparable things. Concretely: the 2026-08-19 account backfill rewrote all 118 curated files in one pass, so in a merged list every one of them would outrank a genuine 08-22 conversation. Each section declares its own orderedBy.

The limit no field can fix. The corpus records what conversations said, never what happened after the newest one. Measured: the newest exchange said "nothing queued, v111 tagged" — 13 commits landed after it. When the answer matters, check the world.

memory({action: "thread", name, forward?, back?})

Read forward from a hit, in sequence. threadLast gives the end of a thread, which is the wrong end of a long one: the resolution to a claim at exchange 200 of a 650-exchange thread is at 201–210. Relevance can't bridge that gap either — the exchange that resolves something often shares almost no vocabulary with the one that raised it ("done", "shipped", "you were right"). Sequence can, and sequence is already in the x-<session>-<ask timestamp> names (e.g. x-fb357616-20260903T054233800Z, which sort as time), so this is arithmetic, not retrieval.

memory({action: "thread", name: "x-fb357616-20260903T054233800Z", forward: 4, back: 1})
  ->  -1  x-fb357616-20260903T053810112Z
      ▶0  x-fb357616-20260903T054233800Z   <- the anchor
      +1  x-fb357616-20260903T060102450Z …          remainingAfter: 30, threadLast: x-fb357616-20260903T235959000Z

offset is relative to the anchor. remainingAfter says how much of the thread the window did not cover, so a long thread stays reachable in one more hop.

memory({action: "verify", name? , text?})

Check a claim against git instead of judging its wording. The corpus records what was said; whether it happened is a question about the world, and for engineering claims the world keeps a record. A cited SHA — the unique fingerprint git gives every commit, 47f71d3 and the like, naming one specific saved change — either exists, landed on the mainline, on a date, touching files, or it does not.

memory({action: "verify", name: "x-df6d25fe-20260818T214812690Z"})
  -> c509e0f [recall-mcp] 2026-08-18  ON MAINLINE  "dream + auto-ingest: a correction signal…"  2 files
     3c1a440 [recall-mcp] 2026-08-18  ON MAINLINE  "auto-ingest: the debounce must run BEFORE the lock"  2 files

latest and thread rows carry verifiedCommits automatically wherever a cited SHA checks out.

• Measured coverage: 404 of 2,319 exchanges (17%) name a real commit. 707 hex-shaped candidates collapse to 355 actual commits, so shape alone proves nothing and every token is checked against git.
• Repos are configured, never inferred — set MEMORY_GIT_REPOS to a :-separated list (; on Windows — the code splits on path.delimiter, so use your platform's). Unconfigured, this stays silent rather than guessing: this server lives in a different repo from the codebase the corpus is about, and guessing would answer confidently about the wrong project.
• onMainline is separate from existence. A commit can sit in the object store after being amended away, or live only on an abandoned branch. "It exists" and "it shipped" are different claims, so both are reported.
• Absence proves nothing. A row with no verifiedCommits cited no SHA.

The reverse join — git → corpus, on time

Every check above reads a commit SHA out of an exchange. That only works when the conversation wrote one down, and measured on a known day it usually doesn't: of 12 commits made during one session, that session's text named 2. The commits happen inside tool calls, while capture records the prose around them — so the identifier normally never appears in the text at all. Reading the corpus harder cannot recover what was never written.

Time can. A conversation has timestamps and so do commits, so the join needs no SHA, no vocabulary and no judgment:

• thread returns commitsDuringWindow — what landed in the configured repos while that stretch of conversation was happening. This turns "I'll commit the fix" — a promise, and the hardest thing in a corpus to resolve — into the record of whether anything actually landed.
• latest returns corpusCurrency — how far behind the world the corpus is, as a count: "N commits have landed since the newest exchange was written." The guidance already says the last word isn't current truth; a sentence is easy to skip and a number isn't.

Evidence, not proof, and labelled that way: a commit inside the window may be unrelated work, and related work can land days later. It narrows "did this ever happen" to "here is what happened at that moment".

Measuring it

npm run analyse-queries   # what callers actually did, and whether retries recovered

eval:state (author's tree only — not in the published package) runs test/state-questions.json, whose answers were written down before the corpus was queried — grading after seeing results produces a test that passes for the wrong reason. Each case also carries a proseControl that is expected to fail; those controls are the measurement behind "query a term filter with identifiers, not prose". Baseline: 6/6 answered, 4 of 6 controls failing as expected. It needs the local store/ corpus, which is gitignored, so it is deliberately not part of npm test.

memory({action: "neighbors", name})

The [[wikilink]] graph, free relevance expansion:

• outbound — links this memory makes
• inbound — backlinks from other memories (verify-protocol has 11)
• unresolvedLinks — [[slugs]] with no matching file
• semantic — top-3 nearest by cosine, which surfaces relatives nobody linked

memory_write({action: "demote", name}) / memory_write({action: "promote", name})

Two-tier mechanics. demote sets metadata.tier: archive in the file's frontmatter, creating a frontmatter block if the file has none. promote removes the line.

Your body text is never deleted or moved — only that one metadata line changes.

The round trip is byte-for-byte reversible only if the file already had frontmatter. If it had none, demoting creates a frontmatter block (including a description synthesised from the body), and promoting afterwards removes the tier line but leaves that block behind. The body is untouched either way. Files with no frontmatter are common enough that this is worth knowing before you demote one.

Hot tier = everything not archived. Anything MEMORY.md lists is hot by definition, so demoting a memory that MEMORY.md still names is refused rather than silently reverted; the response says which.

memory({action: "get", name, outline?, section?, maxChars?, offset?})

A get of the 103 KB build checklist used to blow the MCP output limit outright — so the tool could not read the documents it exists for, and the caller fell back to cat. Three ways in:

Every truncated response carries totalChars, returnedChars and truncated. A slice that looks like a whole document is how a caller concludes something is absent when it is merely past the cut.

Those three fields survive brief: true as well. Brevity may drop provenance; it may never drop the statement of what was left out.

The full response is the right default when you are deciding whether to trust a memory — who wrote it, when, from which account. brief is for when you have already decided to read one and just want the content: most often after a search says the corpus may hold your answer in other words and tells you to open the best weak match. It saves a fixed ~1 KB per call, which is marginal against a long document and most of the response against a one-paragraph note.

Slicing happens after the secrets scrub, so paging cannot reassemble a removed region.

Fenced code is not a heading. This corpus is full of shell snippets whose lines begin # 2. ONLY the intended entries changed…. Read naively those are level-1 headings, and ## MASTER PRE-SHIP GATES returned 426 chars instead of 26,785 — it ended at the first shell comment. The outline went 94 → 50 headings, level-1 count 94 → 1.

memory_write({action: "import", path, dry?, domain?, name?, category?, replace?})

Point it at an absolute path — a file, a folder, or a ChatGPT export — and it brings those memories in. Eighteen formats, no new dependencies:

md markdown txt text log csv tsv json zip rtf rtfd doc docx odt html htm webarchive pdf

textutil (macOS) covers the office and HTML formats and pdftotext covers PDF; archives need nothing — lib/zip.js reads them in-process with node:zlib. A format whose converter is missing is refused by name with the reason — never imported as binary that would poison every search touching it. A ChatGPT export is recognised (conversations.json or its .zip) and its mapping tree is walked in create_time order, because a branched conversation has no single linear list.

converters in the response reports the platform, the probe used (where on Windows, which elsewhere) and what each converter actually is on this machine.

• Refuses any item containing a credential, and names it. A plaintext secret in a corpus is permanent in a way its author rarely intends.
• Never overwrites, so re-running is safe — it reports how many were already imported.
• Skips nothing silently: too-short, refused and unreadable are each counted and named.
• dry: true writes nothing and says so in the future tense: written: 0, plus wouldWrite and wouldWriteNames for what a real run would file (and wouldReplace in place of replaced). The skip counters are unchanged, so a dry run still tells you what it would pass over.
• An argument this tool does not know is refused, by name, on any action that writes — import, capture, index, demote, promote. Read actions stay tolerant.
• It starts the index build itself. An import that wrote at least one file kicks the async index job for the corpus it wrote to and returns indexJobId (+ indexScope); poll it with index_status. Until it finishes the new documents are served from the store rather than ranked. A dry run starts nothing, an import that wrote nothing starts nothing, and a build already running for that corpus is JOINED rather than raced — the response then carries indexAlreadyRunning: true, because that build may have begun before your files landed. MEMORY_IMPORT_AUTOINDEX=0 switches the kick off; the direct read still serves the documents.

Why the import builds the index. Measured over eight import shapes, twice: every imported document was still unindexed 300 s later and would have stayed so indefinitely. Nothing rebuilds a curated index — the 5-minute walker reconciles staging only, and the inline rebuild refuses anything past 8 changed files as a full rebuild in disguise. Retrieval never broke, which is why it went unnoticed: the answers came through the direct store read, a substring scan that answers a token but not a paraphrase and re-reads those files on every query.

Why an archive needs no unzip. The converter probe ran command -v through shell: '/bin/bash', which does not exist on Windows: the spawn failed, every converter read as missing, and .zip import was refused there outright — with the one public check that covers it skipping itself on windows-latest for want of the zip binary to build its fixture. The probe is now where/which spawned with no shell, archives are read in-process, and the fixture is written by writeZipSync, so the check runs on every platform. A stored symlink is still restored and still refused if it points outside the archive, and an entry name that climbs out of the destination is refused by name.

Why a dry run may not say written: 3. It used to. Zero files reached disk and the response said written: 3 with three filenames, because the dry branch reported its intention through the same field a real write uses; only dry: true and one sentence of prose disagreed with the other two signals. A response may not make a claim it cannot back, so the tenses are now separate fields and no field is true of both a dry run and a real import.

Why a write refuses an unknown argument. memory_write({action: 'import', path, dryRun:true}) — a typo for dry — imported for real: the tool is registered with a plain object schema and zod strips unknown keys, so the flag was deleted before the handler ran (measured: written=2, dry=false, curated 19 → 21). The schema now passes unknown keys through so the handler can see them, and a write action refuses one by name, suggesting the argument you meant. A stray key on search still costs nothing and is still tolerated — refusal is for the places where a near-miss costs files.

scripts/import-memories.js is the same thing on the command line, and additionally runs the index and a first dream pass for you.

The corpus knows what kind of corpus it is

This server was built against one software project, and its advice said so: "QUERY WITH IDENTIFIERS, NOT PROSE" was told to every caller. That is measured advice — on a code corpus. Told to someone whose memories are notes for a novel it inverts: they have no SHAs, no flags and no paths, and prose is the only thing they can search with.

Advice now resolves in three layers, most authoritative first:

1. An explicit domain: — code | writing | business | research | planning | prose | mixed.
2. The shape of the query — a SHA, path, CONSTANT_CASE or --flag means technical retrieval whatever the corpus is.
3. The corpus profile — derived once per index by counting structural markers, never by reading a document and classifying it. Reported as corpusProfile with a confidence.

Why both layers. Counting is the floor that needs no cooperation, so a corpus somebody just imported gets sane advice on its first query. But counting can only separate code from not-code: measured across eight corpora, a novel, a business plan, case notes, research notes, recipes and a book about software all score codeScore 0. Only a caller who names the domain can separate those, which is why domain: is first-class rather than a fallback.

The first threshold was codeScore >= 0.35, fitted to this repo's own curated corpus (0.43). It scored 3/8 on pre-registered corpora. Real separation is an order of magnitude lower — prose 0–0.017, code 0.117–0.43. Re-derived on the principle rather than the example: 8/8. test/domain-corpora.json pins all eight.

memory_write({action: "index"}) returns a job

Indexing runs off the request. It used to await buildIndex inline — ~73 s for curated, minutes for staging — so the stale warning told callers to run index and running it returned Error: Request timed out. A tool must never recommend an action it cannot itself complete.

memory_write({action: "index"})                        -> { started: true, jobId }   (~675 ms)
memory({action: "index_status", jobId})          -> { state, indexes, skipped }

One build per index file at a time — a second concurrent index for the same scope reports already being built by job … rather than racing it. wait: true keeps a blocking path for the CLI and tests.

Freshness — an index is a cache of a directory

The incident (2026-08-19). The curated index was last built at 06:18. The corpus files changed at 07:13 and again at 20:46. Every search for the rest of the day answered from the 06:18 snapshot, silently. A session in another chat built a conclusion about the state of the project on top of those snippets — and compounded it by reading each result's modified field, which is the file's mtime at index time, as though it were a live stat.

Two defects: no invalidation rule, and no provenance. Both are fixed.

CHECK. Before answering, search stats the corpus files and compares each one against the mtime the index recorded for it. Exact rather than heuristic: edited, added and deleted files are each detected on their own terms. Measured on this Mac — 122 curated files: 1.00 ms cold, 0.69 ms warm; the 2,104-file staging store: 11.5 ms cold, 8.7 ms warm. Cached for 3 s so a burst pays once.

REPAIR. If anything moved, the existing incremental indexer (lib/index-store.js) runs inline, before the answer. There is one indexer in this repo and the guard calls it; nothing is reimplemented. A one-file edit costs ~3 s end to end (120 files reused, 1 re-embedded).

ADMIT. When the repair cannot be cheap, the query is not blocked. It is answered from the stale index and stamped:

{
  "indexStale": true,
  "indexBuiltAt": "2026-08-20T05:33:03.019Z",
  "staleFiles": 1,
  "staleWarning": "STALE INDEX — these results come from an index built at
     2026-08-20T05:33:03.019Z, and 1 corpus file(s) have changed since:
     1 edited (commit-changes-when-done.md). Not repaired inline because …
     Run memory({action:\"index\"}) before trusting these snippets, and note
     that each result's `modified` is the file's mtime AT INDEX TIME."
}

The repair is refused, by design, when it would not be cheap:

One rebuild at a time per index file (a burst of queries does not start a burst of writers over the same 16 MB), and a failure never fails the search.

Staging is checked and stamped the same way — the mtime comparison is just as cheap there — but it is not repaired inline: its rebuild writes 130 MB in ~14 s and its own ingest hook owns it. Its stamp adds lastIngestAt, from store/.last-ingest.json, because "when did material last arrive" is the more useful question for that corpus.

Which build is answering. Node caches every module at spawn, so an MCP process the client started this morning is still running this morning's code no matter how often the repo is edited — and nothing used to say so. The server now logs its git SHA, branch, pid and start time to stderr at startup, and stamps serverVersion / serverStartedAt on every search response. A running server keeps the old code until the client is restarted (Claude Desktop: full ⌘Q and relaunch; Claude Code: a new session).

Four work corpora + the library, one index each

scope: "all" searches each work corpus against its own statistics and returns them as separate ranked sections under .groups. They are never blended, and that is measured twice:

• Putting 499 auto-ingested exchanges in the curated index cost three probes their answer (22 → 19) and MRR 0.826 → 0.681.

• Putting the 14 handoff documents in the curated index cost MRR 0.8194 → 0.7986 and one absence verdict — while taking zero top-3 slots. Nothing was crowded out. The damage was done entirely by referenceChunks, the corpus-derived p90 chunk count the long-document correction normalises against: 14 long documents moved it 16 → 19, which raised the dense score of every curated memory above 16 chunks and pushed partner-email-rules to 0.3842 against an absence floor of 0.38 — a 0.0042 margin, and the server could no longer say "I have no memory of a Postgres migration".

• Putting 15 other-project memories in the curated index cost MRR 0.8125 → 0.7917 and a rank-1 (measured 2026-08-20; see the next section).

With the handoff documents in their own index, the 32-probe benchmark is bit-identical to the curated-only baseline: MRR 0.8194, absent 4/4, exact 10/10, verbatim 6/6, enum 33/41, and not one rank changed.

The projects corpus — other projects' memory folders

Claude keeps memories per project (~/.claude/projects/<project>/memory), and this server is pointed at one of them. Every other project's folder is discovered automatically, and until 2026-08-20 it was routed into staging by primary: false — so hand-written rules from another project were ranked as though they were raw transcript exchanges (archive tier, no hot boost) and were unreachable at the default scope. Exactly one memory folder exists on this machine, so the defect had never fired. This is the fix before it arms.

They are curated-type content: hot tier, demote/promote allowed, their own account label per file, their project folder carried on every hit. What they do not get is a share of the curated index, and that was measured the same way the handoff corpus was — with a 17-file fixture second project (test/fixtures/projects/…-cli-mcp-server/memory, 15 indexable) pointed at by MEMORY_EXTRA_PROJECT_DIRS:

Two separate damages, worth telling apart:

• Shared statistics — the one that generalises. Not one curated document changed and every probe's score moved (V2 +5.1%, E9 +4.2%, N2 −2.3%): 15 documents joining 122 move BM25's average document length and every idf, so queryIdealScore moves, so the absolute keyword scale moves, so every fused score moves. That is the same mechanism that cost the handoff experiment an absence verdict on a 0.0042 margin. Here it crossed no floor. There is nothing to say it would not next month.
• Competition — the specific one. P2 asks about this project's UI-syntax rule and a memory from another project outranked it. Not by being better; by being in the same ranked list.

Kept separate, the curated index built with the fixture present is bit-for-bit the control: corpusHash 56b48c09…, 122 docs, 1,586 chunks, all three unchanged. The fixture's 15 documents are a 33-chunk index of their own with its own referenceChunks (4, against curated's 16).

Reachable without a scope argument. A memory in its own index cannot be found by a default-scope search, and a standing rule that needs an explicit scope is a standing rule nobody finds. So the advisory router widens to scope: "all" whenever a project corpus exists — the same "widen, never narrow" rule the handoff phrasing uses, and free here because all returns each corpus as its own ranked section. With one memory folder on the machine (today) nothing about routing changes at all.

Read the project field. A rule from another project is a rule about another project. Every search row and every get carries project, account and corpus; project: "this" restricts to the canonical folder.

Day 2 — what happens when a real second project appears

Zero configuration. A session run from another project writes ~/.claude/projects/<other>/memory/foo.md, and:

1. discoverProjectMemoryDirs() finds the folder on the next call — no list to edit, no env var to set. It becomes a projects root, namespaced by the last three dash-segments of its folder name (store/foo.md-style ids, so two projects may hold the same basename).
2. The first search finds the projects index missing and builds it inline before answering — bounded at FRESHNESS.firstBuildMaxFiles (40) files, ~2 s for a 15-document project. Over that bound the search is answered and stamped indexStale with the sentence saying what to run. (Curated at 122 files and staging at 2,100 are both far over the bound, so their behaviour is unchanged.)
3. memory_write({action: "index"}) rebuilds it by default (curated + projects + handoff — the three hand-edited corpora; staging stays opt-in).
4. The router widens, so the new memories are reachable with no scope argument.
5. scripts/auto-ingest.js rebuilds rootsForCorpus('staging'), which no longer contains project roots — so the new folder is not double-ingested as transcript material.
6. Its own MEMORY.md acts as its tier-1 index (inMemoryIndex, the larger hot boost). The bare name MEMORY still resolves to the canonical one — every project has a MEMORY.md, and loadCorpus warns about the shadowing — so ask for the other one by its namespaced id: get({name: "cli-mcp-server/MEMORY"}).

One caveat that is not code: a running MCP server keeps the code and the module state it was spawned with. A client started before this change picks it up only after a restart (Claude Desktop: full ⌘Q; Claude Code: a new session).

The library — category-isolated reference corpora

Books, manuals, policies: imported reference material, which is a different thing from a memory. Daniel's rule (2026-08-26): nothing imported may dilute or even touch work retrieval unless a search names it. Both halves are enforced:

• Isolation by construction. Each immediate subdirectory of each library category is its own corpus with its own .lib-<category>-index.json — own BM25 statistics, own referenceChunks, own profile. The suite's a48 group proves the stronger claim bit-identically: curated corpusHash, every RECALL name and score, and every absence verdict are byte-for-byte the same with library corpora present as with the whole class switched off (MEMORY_LIBRARY=0) — with an absence-probe term planted inside the library fixture the entire time.
• Reach isolation. 'all' stays the four work corpora, and the router never volunteers a category. A category is searched only when named — scope:'books', scope:['all','books'] — or via scope:'everything' (work + every category). Unknown scope names error, listing what exists.
• Read-only (doTier refuses; import's own fs path is the sole writer), archive tier, never rebuilt inline (a changed book is a full re-embed; rebuild with memory_write({action: "index", scope:"<category>"})).
• Import routes and refuses. import with category:'books' files into the category (created on demand). Anything over 200 KB of text, or book-shaped (PDF), without a category is refused before any write — the accident this prevents is a book quietly landing in curated. replace:true supersedes a re-issued document (old version → <category>/archive/, stamped supersededAt, out of the flat scan, never deleted).
• Structure is recovered at import. PDF form feeds become ## p.N page anchors (and running page headers are stripped); docx/html headings become real ##; a plain-text book's CHAPTER lines are promoted (last occurrence of a duplicated designator — a Gutenberg ToC stays plain text). The existing section splitter then chapters the document, so a manual answer cites ts-x73a-user-guide#p-7 — a page a human can open.
• Optional memory-library/<category>/.category.json ({domain, description, note}) declares the category's domain for the advice layer — a statute and a novel are statistically identical prose. Starter categories: books, manuals, policy, legal.
• Validation is pre-registered in test/library-questions.json — bar, grading, and the measured result (invented facts 12/12, manual pages 11/12, absence 10/10, leaks 0; famous-book chapter precision 8/12, honestly short of its 80% bar and recorded as a known limitation).

The handoff corpus

The institutional handoff documents record the state of a phase of work for whoever picks it up next. They lived outside both corpora, so no query could reach them — "what was the state of the corpus refresh" returned the memory summary and never the handoff holding the detail. Daniel approved indexing them (2026-08-19). Fourteen documents, 217 chunks, 2.2 MB.

• Read-only, structurally. readOnly: true travels from the root onto every document, and doTier() — the only writer in the whole tool — refuses. No action can promote, demote, edit or delete one; the test suite asserts the file is byte-identical after both attempts.
• type: "handoff-doc", with the absolute path as provenance (their project is deliberately null — a handoff document is cross-project, and a null project is never filtered out).
• Secrets policy applies unchanged. The pattern guard fired on a bearer token in one of them at index time and redacted it.
• Same staleness guard. These documents change; the check and the inline repair cover them exactly as they cover curated memories.
• Findable by default. They are in their own index, so a default-scope search cannot reach them — which would leave them exactly as unfindable as before. The advisory router therefore widens to scope: "all" on handoff phrasing (handoff, handed over, phase 2, where did we leave, next session, state of the project). It widens; it never narrows, so the curated section is returned untouched alongside.

Graph spread — the [[wiki-links]] finally do something

Every curated memory carries links a person wrote on purpose, and retrieval never read them. After the three legs fuse, each of the top 10 documents now lends alpha (0.15) of its score along its links and backlinks — but only to a document the query already reached, and only if that document clears a similarity gate of its own (0.25). Single hop, computed from pre-spread scores, so a cycle cannot amplify itself. The absence verdict is computed on the pre-spread ranking: spreading reorders, it never answers a refused question.

Measured against a bar fixed before the code existed (test/graph-spread-preregistration.md): curated gold 9/10 → 10/10, MRR 0.850 → 0.950, absence 4/4 and the razor pair unmoved, zero regressions, holding across a plateau of five adjacent grid points. The 10th was "when should I escalate…", which had failed since before the truth-and-recall campaign began: verify-protocol sat at rank 5 while two documents that literally contain [[verify-protocol]] sat above it. ON by default; MEMORY_GRAPH_SPREAD=0 disables. 🟥 alpha 0.30 costs two gold answers — the cliff is one grid step from the default.

Probes — machine-checkable current truth

A memory can record its own check: metadata.probe (a command from the CLOSED eleven-predicate vocabulary in lib/probes.js) plus metadata.probe_expected, compared by equality — arithmetic, never language. The nightly dream pass sweeps them (also memory({action:"probe_status", run:true})), verdicts (FRESH | STALE | UNKNOWN | UNPROVABLE, UNKNOWN-never-STALE on any error) go to the gitignored sidecar .probe-results.json, and memory files are never rewritten. Exact frontmatter grammar + one worked example per predicate: test/PROBE-SYNTAX.md. Dial: MEMORY_PROBE_LEVEL off|cheap|all (default cheap — local file/git/date predicates only; the nightly sweep runs there).

Surfacing (Phase 3b). Twenty claims sampled across the stale-belief taxonomy were hand-adjudicated against reality and written down BEFORE the evaluator ever ran on them (test/probe-calibration.json, bar: ≥12/20 agree and ≤1 false-STALE). The machine agreed on 18/20 with zero false-STALEs, so search results now carry a probeVerdict and the response a probeVerdicts summary. Advisory only, and structurally so: the attachment happens in lib/probe-surface.js, called from the tool boundary after search() has returned — the ranking libraries contain no probe identifier at all, and the suite pins that a STALE verdict leaves the [name, score] list byte-identical. Kill switch MEMORY_PROBE_SURFACE=0 (the sweep keeps running; only the annotation stops).

Proposals (Phase 3c). The nightly dream pass also drafts probes from prose it can already read as a claim — a ship tag beside its sha, a loopback endpoint, an absolute path — and queues them under probe-proposal with the evidence line and the exact frontmatter to paste. It never writes them and it cannot run them: lib/probe-proposals.js imports no evaluator and no process API, and the sweep only reads frontmatter, so an unconfirmed proposal is invisible to it by construction. The rules are narrow on purpose (curated corpus only, loopback/private hosts only, no /tmp, nothing inside a code fence, and the expected value must appear in the prose): the first unrestricted draft produced 1,660 proposals over 659 documents — including a nightly GET at a payment gateway. Today's corpus yields 19 proposals over 11 memories.

Will this touch my memories?

Short answer: it writes frontmatter stamps (tier, modified, provenance), creates new files on import, archives rather than overwrites on import … replace, and never deletes anything from your memory folder. Every one of those writes goes through one door that refuses any edit whose body differs, snapshots the previous bytes to .memory-snapshots/ first, and writes atomically.

If you would rather have the guarantee than the argument, set MEMORY_CURATED_READ_ONLY=1 and the server writes nothing to your memory folder at all — it still indexes, searches, and captures conversations into its own store/. Nothing in retrieval depends on the stamps.

The full inventory, the mutation-tested guards and the recommended setup for imported memories are in MEMORY-SAFETY.md.

Upgrading to 1.6 — exchange names changed on disk

Auto-captured exchanges used to be named by their position in the transcript (x-<session>-0042). 1.6.0 names them by the time the question was asked (x-<session>-20260903T054233800Z). Position was the root of a week of store defects — a changed extractor rule renumbered hundreds of files and left duplicate memories behind, and a deletion bound computed from the ordinal removed a real one. A name that belongs to the exchange cannot do that.

If you have an existing store, migrate once (the server keeps reading either shape in the meantime, and files sort into the same order before and after):

npm run migrate:names            # dry run — prints the plan and every pre-check, writes nothing
npm run migrate:names -- --apply # renames, rewrites name: and Previous:, verifies, refuses on any failure
memory_write({action: "index", scope: "staging"})   # then rebuild the staging index

Back up your store/ first — it is gitignored, so that copy is the only one. The migration refuses to apply unless every file's timestamp compacts cleanly, no two files would share a name, and the new order equals the old order in every session; afterwards it verifies the count is unchanged, every name: equals its filename, no Previous: link dangles and nothing old-shaped remains.

Two related additions: npm run audit:store compares the store against the transcripts it came from (orphans, duplicate bodies, order, dangling links; run it whenever something looks off), and npm run release:capture + npm run install:capture-hooks make the capture hooks run a released copy of the code under dist/capture/ instead of your working tree — so an edit you are still testing can never touch your store on the next hook tick.

The hourly store audit — and the one copy that is not derived

Every other freshness channel this server has trusts something derived. uncapturedSessions compares a transcript to the debounce stamp; captureHealth reads the run log; the recall canary compares the store to the index. When the stamp itself lied — a test fixture wrote store/.last-ingest.json for a live transcript at its full size, so the next real capture decided nothing had grown and skipped — six exchanges were lost and all three said everything was fine.

So once an hour the loaded server spawns scripts/store-audit-tick.mjs, which compares the transcript to the store: it re-runs the real extractor for the 20 most recently touched sessions into a scratch directory and diffs the filenames. That comparison reads the stamp never, which is the point. A missing file younger than the grace window is normal (the exchange being written right now is deferred by design); one older than it is an exchange that had every chance to be written and was not, and the tick repairs it by re-running the extractor with the stamp ignored, under the capture lock, then rebuilds the staging index. Every tick appends one line to store/.ingest-runs.jsonl, and captureHealth — already stamped on every staging answer — now carries both the audit's verdict and the last row of .vanish-report.jsonl (indexed documents that disappeared from disk), which until now nothing read.

The same tick takes a daily snapshot of store/*.md — one gzipped JSONL, {name, mtimeMs, body} a line, in store/.snapshots/store-YYYY-MM-DD.jsonl.gz, keeping fourteen and skipping any day the store has not changed. It exists for one reason: Claude Code prunes transcripts after 30 days by default, and on day 31 the store stops being a derived artefact and becomes the only copy. To put files back:

node scripts/store-restore.mjs store/.snapshots/store-2026-09-05.jsonl.gz --dry    # what it would do
node scripts/store-restore.mjs store/.snapshots/store-2026-09-05.jsonl.gz          # missing files only
node scripts/store-restore.mjs <snapshot> --only x-b58a69af-20260905T044521647Z.md # just one
node scripts/store-restore.mjs <snapshot> --force                                  # also overwrite

It restores missing files only unless you ask for --force: the reason to run it is that something removed files, and overwriting the ones still present would roll the store back to the snapshot — turning a partial loss into a total one. MEMORY_STORE_SNAPSHOT_HOURS=0 turns snapshots off; MEMORY_STORE_SNAPSHOT_KEEP changes how many are kept.

Versioning the memory folder

The curated memories had no version control, so a bad overwrite was unrecoverable — and the folder already held two hand-made .bak files someone created because there was no other way to undo a change. scripts/commit-memories.js gives it a history:

npm run memories-status    # what has changed since the last commit
npm run commit-memories    # commit it now (the Stop hook does this automatically)

It runs from the Stop hook as its own entry, not appended to the ingest command — if they shared a shell line, a git failure would take memory capture down with it, and capture matters more than versioning. At most one commit per turn, only when something changed, and every path exits 0 so a hook can never fail a turn.

To recover a clobbered memory:

git -C "$(node -e 'import("./lib/config.js").then(m=>console.log(m.memoryDir()))')" log --oneline -- some-memory.md

then git show <sha>:some-memory.md.

Local only, deliberately. A memory corpus tends to accumulate credentials — an SSH password pasted into a note, a token in a runbook. On disk that is a pre-existing fact you can fix; in a pushed history it is permanent and off-machine, surviving any later deletion unless the history is rewritten. The script therefore never adds a remote, never pushes, and refuses to run if a remote is configured while an unmistakable secret is still present, naming the offending files.

"Unmistakable" is deliberately narrower than the redaction vocabulary, because this decision blocks you from versioning your own notes and a false positive there is expensive. Three shapes block a commit: sshpass -p '…', a -----BEGIN … PRIVATE KEY----- block, and an AWS AKIA… key. A line like password: hunter2 is redacted everywhere it could be served — the index, search results, get — but does not block a local commit. The asymmetry with import, which refuses such a file, is intentional: declining to copy a file in is cheap and tells you which one, whereas declining to record your own history is not. Offsite backup is a separate decision that needs the credentials moved out first.

Secrets policy

Assume the corpus will contain plaintext credentials sooner or later. Four mechanisms, each enforced at index time and again at output time:

1. Filename denylist — secrets-exclude.json → excludeFiles. A file listed there is never indexed; get and neighbors return a refusal. Ships empty: add your own, and note that the list itself is public, so name files by path rather than by what they contain.
2. Frontmatter opt-out — metadata.secret: true gets the same treatment. Re-checked on every get, so marking a file secret takes effect immediately, before any rebuild.
3. Section scrub — for the case where one section of an otherwise useful file is the problem. Name the file and its heading in sectionScrub and the file is indexed with that section stripped, with get returning the scrubbed version. Configured per machine; ships empty.
4. Pattern guard (backstop) — password-shaped text is redacted to [REDACTED:<class>] before the index file is written and before any tool response leaves the process, with a WARN naming the stage it fired at. The class names the pattern that fired (credential-shaped, token-shaped, key-shaped, known-credential — a CLOSED vocabulary; an unlisted class fails the config load) and never encodes anything about the redacted content. Older corpus text still carries the bare [REDACTED] form; both generations are inert to re-redaction.

Mechanism 4 exists because curated lists go stale. It currently catches six real chunks drawn from two large memories in the author's corpus — files nobody thought of as credential-bearing.

No plaintext credential lives in this repo. secrets-exclude.json stores known literals as sha256 hashes of their lowercased form, and the tests detect leaks the same way. A test that hard-coded the password to grep for would itself be the leak.

If secrets-exclude.json is unreadable, the server fails closed: every file is treated as excluded rather than risk indexing an unfiltered corpus.

A password said in a sentence

The shapes above are machine shapes — a header, an assignment, a prefixed key. The shape a person uses is a sentence, and until 1.7.3 the guard had no rule for it: "My test password is Hunter2-Xk9!pass and my key is AKIAIOSFODNN7EXAMPLE" stored the key as AKIAIOSF•••••••••••• and the password in clear, in the memory's description and in its search/latest snippet.

password-in-prose covers it, and stays deliberately narrow. It fires only when a trigger word (password, passwd, passphrase, passcode, pin) is followed by a link (is, was, :, =, ->, set to, changed to) and the value earns it: quoted (any shape — a quote is how a person says this is the value), or 8+ characters carrying a non-letter, or 12+ characters. A bare lowercase word never counts: measured over 2,952 real exchanges it scored 0 true and 3 false ("provided", "supposed", "extended"), so "the password is required" is left alone, and so is **provided** in bold. A trigger inside an inline code span (`Password:`) is not a link to whatever follows the closing backtick. Teams' Passcode: <value> lines are covered by the same rule.

The rule also has to survive the guard's last sweep, which runs over the serialized index — JSON, where every quote in the corpus is \". The first cut let a backslash be a value character, so must not redact \"password is required\" matched (required\ is nine characters carrying a "non-letter"), the escape was eaten, and buildIndex correctly refused to write an index that no longer parsed: npm run index exited 1 and the staging index could not be rebuilt at all. A value character is therefore never \ " ' ` , ; . ) } ] > or whitespace, and the non-letter that qualifies an unquoted 8+ token must be a digit or one of !@#$%^&*_+-=~?/| — never a backslash, a quote or a bracket. (a94) now round-trips every fixture through JSON.stringify and builds a real index over a corpus that holds an escaped quote.

A note on -p

An early version of the guard redacted any quoted value after -p and shredded 106 chunks — unzip -p "$ZIP", mkdir -p "...". The rule is now scoped to commands that actually take a password there (sshpass, mysql, psql, smbclient, …). A secrets guard that mangles the corpus gets turned off, which is the real failure.

The embedding contract

Xenova/bge-small-en-v1.5, quantized ONNX, 384-dim, mean pooling, L2 normalised. bge is asymmetric: the prefix Represent this sentence for searching relevant passages: goes on queries only; passages are embedded bare. Getting that backwards costs recall silently — no error, just worse answers forever.

Because vectors are meaningless without the recipe that produced them, .memory-index.json opens with a self-describing header:

{ "formatVersion": 1, "model": "...", "queryPrefix": "...", "pooling": "mean",
  "normalize": true, "dim": 384, "chunkWords": 200, "chunkOverlapWords": 40,
  "chunkCount": 1510, "docCount": 110, "corpusHash": "...", "builtAt": "..." }

The loader compares every contract field against the running configuration and refuses the dense half on any mismatch — logging which field disagreed — then serves BM25-only. It never silently returns vectors built by a different recipe. Same for an unparseable index: mode: "unavailable", no crash.

If @xenova/transformers cannot load at all (not installed, model not cached, no network on first run), everything degrades to BM25-only loudly: an ERROR log with the fix, and mode / degradedReason on every search response.

That last paragraph — the model failing to load at search time — is the one claim here with no test behind it, and it is untested deliberately. The state it describes is "a dense index on disk, and no model to embed the query with", and this codebase cannot reach that state: lib/index-store.js:266 refuses to build an index at all when the model is unavailable (Daniel's 2026-08-30 rule — a stale dense index beats a fresh keyword-only one — with deliberately no env override, because an escape hatch is how "temporarily" becomes permanent). Constructing the state in a test therefore means adding exactly the override that rule exists to forbid, so the branch is left unexercised and said so here instead. Note the scope: mode: "bm25-only" is also reached when the index header disagrees with the running contract, and that path is tested (test/run-tests.js:1433, a deliberately corrupted header). Only the model-unavailable branch is unproven. What you would see if it fires: one startup ERROR line reading DENSE RETRIEVAL DISABLED — embedding model unavailable: … followed by the fix (re-run npm run index with network access so the model caches), mode: "bm25-only" and a degradedReason on every search response thereafter, and rankings that still carry keyword and phrase evidence but have lost semantic matching entirely — a paraphrase will no longer find a memory that shares no words with it.

Model cache

./.model-cache (gitignored, ~35 MB). Populated on first npm run index with network access. If the machine has the same model cached elsewhere, copying Xenova/bge-small-en-v1.5/ into .model-cache/ works too — it is a plain directory of files.

Registration

Claude Desktop

~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "memory": {
      "command": "npx",
      "args": ["-y", "agentic-recall@2"]
    }
  }
}

If the file already has mcpServers, add just the "memory" block inside it rather than replacing what is there.

🟥 QUIT Claude Desktop completely BEFORE editing its config; edit; then launch; then verify serverVersion. Claude Desktop holds this file in memory and writes its own cached copy back over it when it quits, so an edit made while the app is running is discarded the moment it exits — with no error, and a file on disk that looks untouched. Measured on a real machine (2026-09-06): a config edited to 1.7.2 while Claude ran came back as 1.6.3 after the next quit-and-relaunch, and the file had reverted to its previous contents. Two releases were installed as files and never actually loaded because of it. Quit with ⌘Q on macOS, or tray icon → Quit on Windows — closing the window only hides it.

Claude Desktop also caches MCP tool schemas at connection time. After the relaunch, start a new conversation before the memory tool appears; a conversation that was already open will not see it.

Then verify. Every response the memory server returns carries a serverVersion field. If it names an older version than the one you installed, the edit was reverted — quit completely and do it again. That field is the only way to tell a saved config from a discarded one.

Claude Code

claude mcp add memory --scope user -- npx -y agentic-recall@2

Existing Claude Code sessions pick it up on the next session, not the current one.

🟥 Capture needs Claude Code's transcripts — retrieval does not

Retrieval works everywhere. Point MEMORY_DIR at any folder of markdown and search, absence verdicts, latest, get and the git-verification layer all work — with any MCP client, on any platform. That is the whole product as far as searching goes, and it depends on nothing but your files.

Capture is narrower, and it is worth knowing before you rely on it. It has nothing to do with Claude's own memory feature — it reads the .jsonl transcript files Claude Code writes for every session automatically (the same ones --resume and --continue use). No opt-in, no setting. But only Claude Code writes them where this looks:

Why Desktop chat cannot be captured: those conversations are not written to disk as transcripts at all. Measured on macOS — Desktop's entire local storage is a few hundred KB of browser state, far too small to hold chat history, and the only .jsonl files under its application-support directory belong to agent-mode sessions. The conversations live in your account, not on the disk this server can read.

So if you install this on Claude Desktop and expect your chats to start appearing in staging, they will not, and nothing will tell you so. Retrieval over the notes you write is unaffected.

That is the whole of capture setup

A loaded server keeps time. While the connector is switched on, the server spawns a capture walk every five minutes — the same scripts/timed-capture.mjs you can run by hand with npm run capture. Nothing else is required for conversations to be remembered: on Windows, nothing else at all.

That matters because the two older triggers each miss something, and one of them is macOS-only:

The hook and the LaunchAgent are now belt and braces, not requirements. Keep them if you have them (a hook captures the moment a turn ends rather than up to five minutes later); do not install them if you do not.

The exchange being written right now. A timed walk deliberately leaves the last exchange of a live turn alone: it is still growing, and writing it every five minutes would re-embed it every five minutes and make a truncated answer briefly searchable as though it were finished. A hook does not have that problem — a turn that has ended is finished by definition — which is why the hook never defers. With no hook installed, though, "the next pass will get it" was never true for a chat that simply stopped: the last exchange waited for the hourly store audit (measured on a Windows install: ~14 minutes, up to ~75). So the timer now asks whether the turn is still ALIVE. A transcript that has not moved for MEMORY_INFLIGHT_QUIET_MIN minutes (default 10) is over or abandoned, and its last exchange is captured exactly as the hook would capture it — stamped inFlight: true, and rewritten in place with the flag cleared if the turn ever resumes and ends.

And most of the time it does not have to wait at all. Every assistant record in the transcript carries stop_reason, and end_turn (or stop_sequence) means the model STOPPED — the same event the Stop hook fires on. So a timed walk captures that exchange on the very next tick, with no quiet threshold and no inFlight flag, because there is nothing provisional about it. Only a turn whose last record is tool_use (still working) or absent (an older transcript shape) is deferred, and the quiet rule above is what eventually releases it. Measured with a live server at a 2-second interval: 1.7 seconds to store, against three ticks and eleven minutes that never captured it at all.

The same field is why inFlight: true now means something. It used to appear on every hook capture ever made — at Stop time there is by definition no next user turn, which was the only test — so a flag meant to warn "you may be reading a draft" was permanently on. It is now stamped only when the transcript itself does not say the turn ended.

Not every session is a conversation. A <scheduled-task …> transcript is a robot run, and capture refuses it by design. That refusal is now one shared predicate (lib/capturable.js), so the writer, the walker and uncapturedSessions agree about it: such a session is never queued for capture, never counted as a gap waiting to be filled, and never named in the "captured by the next timer tick" promise — it is summarised as excludedSessions: {count, reasons} beside the count instead. Before this, four of them sat permanently in a count of eleven and took four of the walker's eight slots on every tick (measured again while fixing it: 10 uncaptured became 0, all ten scheduled tasks). MEMORY_CAPTURE_INCLUDE_TASKS=1 opts them back in.

A few knobs, all optional:

Several servers is the normal case — Claude Desktop runs one, and Claude Code runs one per chat — so the timer is built for it. Every walk takes a lock (store/.timed-capture.lock) and records when it started (store/.timed-capture-last.json), whoever launched it. A server whose neighbour, or the LaunchAgent, walked forty seconds ago simply does not fire; if two do fire together, one walks and the other exits saying so. The jitter keeps N servers that booted at the same moment from all reaching the same conclusion in the same millisecond.

The server never writes to the store itself. It spawns the walker, which spawns the same per-session writer the hook uses, so there is exactly one write path with one lock — see the header of lib/scheduler.js.

Layout

index.js                 MCP server entry (stdio; stderr-only logging)
tools/memory.js          the one gateway tool, dispatching on `action`
lib/config.js            paths + THE EMBEDDING CONTRACT + retrieval knobs
lib/corpus.js            frontmatter parse, headings, wikilinks, tier read/write
lib/bm25.js              tokeniser (light stemmer) + Okapi BM25F, 3 field groups
lib/lexical.js           the phrase leg: best window + windowed snippets
lib/embed.js             @xenova loader, chunking/unchunking, query asymmetry
lib/index-store.js       build / validate-header / load .memory-index.json
lib/freshness.js         the staleness guard: stat pass, comparison, inline repair
lib/version.js           which git SHA the running process was spawned from
lib/search.js            fusion, long-doc correction, absence verdict, provenance
lib/secrets.js           the four exclusion mechanisms
secrets-exclude.json     denylist + scrub config + hashed known literals
scripts/build-index.js       npm run index
scripts/verify-stdio.js      npm run verify — raw JSON-RPC, no client needed
scripts/probes.json          the 32-probe benchmark set (queries are verbatim)
scripts/verify-stdio.js      npm test — drives the server over raw stdio
scripts/measure-*.js         where each tuning constant came from
scripts/ingest-transcript.js a conversation becomes exchanges (x-<session>-<ask time>); folds mid-turn
                             messages and subagent reports into the exchange they belong to
scripts/auto-ingest.js       the Stop-hook entry: lock, debounce, run log, staging reindex
lib/scheduler.js         the in-server timer: spawns the walker, writes nothing itself
lib/fs-retry.js          rename-with-retry — a Windows swap fails while anything holds the file
scripts/timed-capture.mjs    npm run capture — walks every active transcript on a timer
scripts/migrate-stable-names.mjs  npm run migrate:names — one-time move off positional names
scripts/audit-store.mjs      npm run audit:store — store vs transcripts (lib/store-audit.js)
lib/store-audit-tick.js  the hourly audit: classify, alarm, self-heal under the capture lock
lib/store-snapshot.js    the daily gzipped-JSONL copy of store/*.md, keep 14
scripts/store-audit-tick.mjs  one audit tick — spawned by the in-server scheduler
scripts/store-restore.mjs     put store files back from a snapshot (missing only, unless --force)
scripts/release-capture.sh   npm run release:capture — the copy the hooks actually run (dist/capture)
scripts/install-capture-hooks.sh  npm run install:capture-hooks — point hooks + LaunchAgent at it
test/run-tests.js        npm test — exit code is the verdict
test/fixtures/projects/  a FIXTURE second project's memory folder (17 hand-written
                         files). Deliberately NOT under ~/.claude/projects, so it can
                         never be mistaken for a real one; reached only by
                         MEMORY_EXTRA_PROJECT_DIRS. It is what made the
                         other-projects routing measurable with one folder on the
                         machine.

What it writes down about you

Local-only, and worth knowing before you point this at anything sensitive.

Every query is logged verbatim. .query-log.jsonl in the server directory records, per search: the query text, the scope, the top result's name, the confidence, and whether it refused. It exists so the retrieval work can be measured against real questions rather than invented ones, and npm run analyse-queries reads it. It is gitignored, never leaves the machine, and no part of it is sent anywhere.

Turn it off with MEMORY_QUERY_LOG=0, or point it elsewhere with a path. Nothing else changes if you do — it is diagnostics, not a dependency.

The other files the server writes beside itself, all gitignored: the index (.memory-index.json and friends), the vector cache, the probe sidecar, and the curation state. All of them mirror corpus text, which is why none of them is ever committed and why scripts/commit-memories.js refuses to add a remote.

Known limitations

Measured, reproducible, and not yet fixed. Listed here because finding them yourself and not seeing them mentioned is worse than reading about them. And one thing that is untested by design rather than unmeasured: the BM25-only fallback for a model that fails to load at search time — see The embedding contract.

A long verbatim quote can retrieve worse than a short one. Quote a sentence from a memory's body and the memory usually comes back first. Keep adding words from that same sentence and it can drop out of the results entirely:

Why: the keyword leg drops any document scoring below covFloor × ideal, where ideal assumes the query's terms at full field weight. Body text carries fieldWeights.body = 0.3 against name 3.0, so a body-only match sits near 30% of ideal and cannot clear a 60% floor. It is then carried by the dense and phrase legs alone, and long queries full of common words dilute both.

Lowering that floor was tried and rejected on measurement — over 45 verbatim body sentences it moved nothing (rank-1 25 → 25, missing 18 → 18), so the floor is not the binding constraint. The same measurement showed the misses concentrate in large documents whose #section children compete with their parent: restricted to whole documents the leg is healthy (rank-1 24 of 28).

Practical workaround: quote a short distinctive fragment rather than a whole sentence, or use latest with an identifier — that path is a literal term filter and finds body content the ranker misses, including commit SHAs.

CR-only line endings (classic Mac) do not parse as frontmatter. CRLF and a UTF-8 BOM both parse correctly. A CR-only file is reported honestly (hasFrontmatter: false) rather than silently mis-parsed, but its name, type and headings are lost. Convert such files to LF or CRLF.

What it costs at size

Measured on one laptop, so treat them as shape rather than benchmark. The fixed cost is the embedding model (~215 MB resident); everything above that scales with chunks, not documents.

Three things worth knowing before you point this at something large:

• Build time follows your biggest file, not your corpus. 600 ordinary notes index in about 90 seconds; a single 4.6 MB document takes 163 seconds on its own. If a rebuild is slow, one file is usually the reason, and the build now names it.
• Vectors are Float32Array in memory and base64 float32 on disk (index format v2; lib/vec.js is the single representation authority). That is 1,536 bytes per 384-dimension vector against ~3,700 for the plain JavaScript array this used to keep — measured, not estimated. It costs nothing in accuracy: the embedding model emits float32, so float64 stored no extra information, and the largest cosine difference between the two over 200 vector pairs is 4.9 × 10⁻⁹, against the 10⁻³–10⁻² score gaps that actually decide a rank. What now dominates resident memory is the parsed index itself — chunk text, names, descriptions and the BM25 postings — not the vectors: 17,815 chunks hold only ~27 MB of vector data inside an 812 MB process.
• Search stays fast: 32 ms at 600 documents, ~100 ms at 2,790. It is the memory, not the latency, that will bother you first — the ceiling is your machine's RAM, and one process holding the index is what sits in it.

Nothing here is a hard limit; they are the numbers, so you can decide.

Environment overrides

Section documents (MEMORY_SECTION_DOCS) — on by default

project-changelog is 635 KB of ~40 version entries and release-build-checklist is 100 KB across 21 sections. The useful unit is one section, and RETRIEVAL.longDoc penalises a document by its chunk count, so neither could win a query about its own content.

Large memories (>= 20 KB, >= 3 ## sections, never exchanges) are indexed as parent#section children; the parent becomes a small navigation stub. Measured against questions registered before any of them was run (test/section-questions.json):

Recall, MRR and the max-over-chunks artefact are identical - seven questions that returned nothing useful now return the exact section. Baseline 0/12 understates it: only 4 of the 12 returned even the parent document. npm test is 512 passed / 2 failed in both arms (the same two).

Three things had to be true, and none was what the plan predicted. The long-document penalty was never the constraint - MEMORY_SECTION_BETA and MEMORY_SECTION_WAIVER barely move the result across the whole grid. What mattered was:

1. IDF counts documents, not index rows (docFile, lib/bm25.js). 138 changelog sections inflated missingIdf = idfOf(0), pushing orphanShare over its floor so the absence guard withheld a correct answer. The scorer and the normaliser must share that basis, or a section whose terms concentrate in one file has its keyword score collapse.
2. A child owns its identity - its own hash, and only its own heading. Inheriting the parent's hash into a file-keyed reuse map handed every child the parent's chunks and crashed the build with RangeError: Invalid string length; inheriting headings gave each child a claim on all 275.
3. The stub reproduces nothing. Headings are already a separately weighted field and each is the first line of its own child, so repeating them made an 18 KB keyword magnet that answered "what ports are used across all these projects" with the changelog's table of contents.

Superseded version sections lose their boost. The changelog is 135 sections at hot tier where the newest two or three matter operationally. Sections whose heading carries a version are ordered arithmetically and all but the newest MEMORY_SECTION_KEEP_VERSIONS are demoted to archive — searchable, get-able, just not boosted (18 hot, 117 archive). A section with no version is left alone: unorderable is not the same as old. The only language rule is a literal ⚠️ STALE marker, which is why "Helper scripts — ⚠️ STALE, do not trust" is demoted while "Counts in this file go stale — VERIFY the count" is not; the second is advice about staleness. Nothing here reads a section and judges it.

MEMORY.md is never split, at any size. It is the index — a list of pointers whose parts mean nothing apart — and it is loaded into context every session.

Changing the setting changes the corpus hash, so the index is detected stale automatically. A full rebuild is not an inline one, so the first search after switching may report indexStale and ask for memory_write({action: "index"}).