NexusMem

Your coding agent can read git log. It cannot read the four things you tried last Tuesday that didn't work.

NexusMem records what actually happened on your machine (shell commands and their exit codes, git history down to the patch of each changed file, project docs, optionally your assistant transcripts) into a local SQLite database, and serves back a ranked, token-budgeted slice of it on demand. Everything stays on disk. No account, no cloud, no telemetry.

The shell history is the part worth caring about. Git tells an agent what shipped. Shell history tells it what was attempted, in what order, and which commands exited non-zero. That information exists nowhere else, and it disappears when your terminal scrollback rolls over.

Contents: Try it · Exact shell capture · Failure → fix chains · How retrieval works · Session summaries · GitHub issues & PRs · Use it from an agent · What it costs you · Staleness & provenance · Where it breaks · Commands · Cross-project recall · On disk · Development

Try it

From inside any git repository:

npx nexusmem init
npx nexusmem sync

Then ask it something. Real output from this repository, top 2 of 5 hits:

$ nexusmem query "windows spawn failure"

Relevant history for: windows spawn failure

- 2026-08-09 [observed] fix: distinguish a failed git spawn from "not a git repository"
  readRepoInfo collapsed three unrelated failures into one error: git running and reporting
  the path is not a work tree, git not being installed, and the process failing to spawn at
  all. Dogfooding hit the third case in two separate sessions...
- 2026-08-09 [authored] README.md — Before a tagged release
  - [ ] Retry on transient process-spawn failures on Windows

[observed]/[authored] is the provenance tag (see Staleness & provenance) — a commit is a directly observed event, a doc section is a written claim that could go stale.

A commit and a docs section, ranked against each other, inside whatever token budget you gave it. Nothing was summarized by a model on the way out; the ranker just decided what not to send. (One optional source, session summaries, does run a local model — but at ingest time, never on the way out. What you query is always stored text.)

For a sense of what actually accumulates, here is nexusmem status on this repo after two days:

527 node(s)  2026-08-08 .. 2026-08-09
       321  shell_command
       130  conversation_turn
        60  doc_section
        16  git_commit

Sixteen commits. Three hundred and twenty-one shell commands. The commits were already retrievable by any agent with a terminal. The rest was not.

That conversation_turn row only appears because this corpus was synced with --conversation. Assistant transcripts are the one source that is off by default and stays off until you opt in, since they are the likeliest place for a pasted credential to be sitting. A default install indexes git commits, their diffs, shell and docs.

Requirements: Node 22 or newer, and git. Node 20 will not work, because better-sqlite3 ships no prebuilt binary for it and Node 20 went end-of-life in April 2026. Ollama is optional and only affects semantic search (see below).

Optional: exact shell capture

Scraped history files (PSReadLine, .bash_history, .zsh_history) give you command text and not much else. The hook gives you working directory, exit code and a real timestamp:

nexusmem hook install

It wraps your existing PowerShell prompt rather than replacing it, is idempotent, and nexusmem hook remove undoes it cleanly.

Exit codes are what make this worth installing. A failed command is a stronger signal than a successful one, and without the hook there is no way to tell them apart.

Failure → fix chains (opt-in)

nexusmem sync --link-failures

After a normal sync, this walks every failed shell_command (non-zero exit code) and looks for whatever later resolved it, using two independent heuristics: a later command in the same project and working directory, exact same normalized text, that exited 0 within 24h (same-command retry); and, separately, the best full-text match among nearby conversation turns or session summaries, requiring every significant word of the failing command to appear, not just one (conversation bridge). A failure can be linked by either, both, or neither.

Both links are surfaced in query results. The conversation-bridge heuristic originally matched on any shared word, and dogfooding against this repo's own real history found it wrong on roughly half its links — a shared word as generic as "npm" was enough to link an unrelated discussion. Requiring every significant word fixed that: re-dogfooded against the same corpus, every resulting link (the full set produced, not a sample) checked out correct on manual review of the full text, not just the summary.

When a linked failure appears in a result set, its fix rides along immediately after it, inheriting the failure's own relevance score rather than needing to match the query on its own merits. That is the point: a query about why something failed shouldn't need to separately guess the words used in whatever fixed it. This works across projects too — query --all-projects chains a failure to its fix using whichever project's own database recorded the link, since links are always local to the project they were found in.

$ nexusmem query "why did npm whoami fail"

- 2026-08-12 [observed] shell: npm whoami  (exit 1)
- 2026-08-12 [observed] shell: npm login   (exit 0)  -- linked as the fix

How retrieval works

Every source normalizes to the same MemoryNode shape, so a commit, a shell command and a docs section compete on equal terms. Retrieval runs BM25 over FTS5 and, if an embedding model is reachable, a vector search over sqlite-vec, fused with Reciprocal Rank Fusion on rank position only, never raw scores — a BM25 cost and a vector distance live on unrelated, unbounded scales, and position is the only thing they agree on.

Ranking then multiplies three factors:

score = relevance × signal^0.215 × recency^0.288

relevance comes from the query. signal (a fix: commit outranks a chore:; a failed command outranks a successful one) and recency are priors that hold before any query exists. Each factor is floored into [floor, 1] rather than [0, 1], so one weak dimension can't zero out a strong match.

The exponents bound how far signal and recency, together, may overturn relevance: at most a 2× gap across their whole range, applied jointly rather than per-prior. That's deliberate — the score multiplies the two priors, so capping each at 2× separately still let the pair overturn 4×, and that hit hardest on fresh, high-signal commits made during an active working day. The bug that exposed this: two unrelated same-day fix: commits outranked the docs section that actually answered the query. See retrieval/rank.ts for the full derivation.

Without Ollama, vector search is skipped and you get BM25 only — fully supported, not a degraded state; sync and query both succeed and simply do less.

Session summaries (optional, local model)

With sources.session.enabled, each finished session becomes one distilled node next to the raw exchanges — what was decided and why, rather than forty individual turns. It runs a local Ollama chat model (qwen2.5:3b by default); nothing is downloaded automatically and nothing leaves the machine.

nexusmem scan-session --dry-run

That prints the exact prompt a session would produce, after redaction and budget trimming, without calling the model.

Three things bound the cost. A session is only summarized once it has been quiet for settleMinutes (default 30), so a session in progress is not re-summarized on every sync. The prompt is hashed, and an unchanged hash skips the model entirely — on this repo a steady-state sync of 14 summarized sessions takes 0.25s and makes no model calls. And maxSessions (default 10) caps how many reach the model per run; the rest are reported as queued and picked up next sync.

What it is actually like, measured on 14 real sessions with qwen2.5:3b. The summaries themselves are good: decisions with their reasons, in the shape the prompt asks for. Titles are less reliable — the model returned a usable one about a third of the time, and otherwise produced a conversational preamble, a stray bullet, or a bare "Summary of the Session". Those are rejected and the title falls back to the first line of the question that opened the session, which is always specific even when it is not elegant. Compliance was worst on long sessions and on transcripts not in English. A larger model (qwen2.5:7b) is the lever if the titles matter to you; set sources.session.model.

GitHub issues & PRs (optional)

With sources.github.enabled, each issue and PR on this repo's github.com remote becomes one node (title, opening post and every comment, folded together) alongside the raw discourse a discussion already leaves in shell/conversation history. Off by default — not for a sensitivity reason, but because it's the first source with a real external dependency: it reads via the gh CLI, so it needs gh installed and authenticated (gh auth login), and it makes live network calls instead of only reading what's already on disk. A repo with no github.com remote, or an unauthenticated gh, is a silent no-op either way.

nexusmem scan-github

previews the nodes a sync would produce, same as the other scan-* commands. maxThreads (default 100) and maxCommentsPerThread (default 100) bound one sync's cost; since is tracked as its own cursor, so a repeat sync only re-reads threads that changed. Dogfooded against this repo's own 14 real issues/PRs: ingest took under a second, and a query for "labelled retrieval regression corpus" correctly ranked issue #8 — the one that asked for it — first.

Use it from an agent

{
  "mcpServers": {
    "nexusmem": {
      "command": "npx",
      "args": ["-y", "nexusmem", "mcp"]
    }
  }
}

Three tools over stdio: search_memory returns the packed context block, sync_project ingests, and get_status reports what is currently remembered. Each takes an explicit projectRoot, because an MCP tool call carries no shell working directory. sync_project runs init for you if the repository has not been set up.

What it costs you

Two numbers get conflated in tools like this, so they are kept apart here.

Packer efficiency is how much the ranker trims from its own candidate set. On this repository's corpus it runs 81–84%. It is useful for tuning the ranker and useless as a claim about your bill, because the baseline is hypothetical: without NexusMem those candidates were never going into your context window in the first place.

End-to-end saving compares the packed context NexusMem actually sends against reading, in full, the same files its own ranking identified as relevant to the query. Measured with scripts/benchmark.ts (npm run bench), which anyone who clones this repo and points it at a synced corpus can re-run from scratch:

Both clear the original >70% target ("cut API token spend versus sending full context"), and the vite run is the first measurement at the scale that target was always described as applying to.

Read the methodology before quoting either number — it's a narrower claim than it looks:

• Graded against NexusMem's own ranking, not an outside answer key: the file set is whichever files the packed nodes for that query touch. This measures what the pack step saves once retrieval already picked a candidate set; it doesn't independently verify that set was the right one.
• Query sets are mechanical, not cherry-picked (see scripts/benchmark.ts): vite's is an even sample of well-explained fix/feat/perf/refactor commits plus rationale-bearing doc headings; this repo's reuses real historical prompts verbatim, several of which are broad task instructions rather than narrow questions — part of why its number sits below vite's.
• git log -p baselines can be enormous — one vite query's baseline hit 7.5M tokens because a file in its resolved set has that much history. At that scale, "just read the file's history instead" stops being a viable alternative at all.
• Supersedes the old ~40% figure, which was hand-tallied from two hand-picked queries against this repo alone with an unstated baseline. Not wrong, just underspecified — this replaces it with a stated method and a script that reproduces it.

One thing that is not a percentage: shell commands and conversation turns have no cheap grep equivalent. Without something recording them, they are gone, not merely more expensive to find.

For how these numbers compare to a similar tool's own claims, see docs/competitor-comparison.md (vs. projectmem) and docs/competitor-comparison-yesmem.md (vs. YesMem, including native Windows support vs. its documented WSL2 requirement).

Latency on a ~530-node corpus, warm, p50 over 10 runs:

All the SQLite work totals about 5 ms. The embedding call is the only thing on this path worth optimizing, and it is somebody else's process.

Staleness & provenance

Two things a memory layer needs and this one only partly has: a way to tell an observed fact from a guess, and a way to retire a conclusion once something contradicts it. This section is what exists and what doesn't.

Every node carries a provenance, a four-tier trust hierarchy set once per collector at ingest time: observed (a commit that landed, a shell command's real exit code) > authored (a doc section — a human's own written claim) > recorded (a conversation turn — verbatim, but talk about events rather than the events) > derived (a session summary — a model's distillation). The tier is shown as a tag on every query result and decays retrieval weight — the lower the trust, the faster a node fades from ranking as it ages. The ordering is the design claim; the exact decay ratios are judgment calls, not measured optima.

nexusmem stale

Lists non-observed nodes old enough (45+ days by default) that nothing has confirmed they still hold — a heuristic on age and provenance, not on content. It writes nothing; you decide which candidates are actually wrong. Any candidate the SLM has already flagged (see below) is decorated with its standing likely superseded by suggestion — reading those costs nothing, so the plain command stays instant and offline.

nexusmem mark-stale <oldNodeId> --supersedes <newNodeId>

Links newNodeId as the replacement for oldNodeId. The ranker down-weights the old node from then on (it stays queryable, just usually loses to its replacement) — nothing is deleted, unlike forget.

nexusmem stale --check-contradictions

For each candidate, finds the most similar newer node (local embedding search) and asks a local SLM (Ollama, qwen2.5:3b by default) whether it actually contradicts the older one — real content comparison, not just age. A match is printed as likely superseded by <id> <title> -- <reason> under the candidate. Every judgment (either verdict) is memoized, so a judged pair is never sent to the model again; nothing else is written — supersedes stays yours to set via mark-stale.

This also runs automatically during sync — at most 3 new judgments per run (configurable via the contradictions block in .nexusmem/config.json; set autoCheck: false to turn it off), only when the embedding provider was reachable anyway, and free on repeat syncs thanks to the memoization. New and open suggestions show up in the sync summary, nexusmem status (a flagged line), and plain nexusmem stale.

What this doesn't do: it is one small model's yes/no judgment on one older/newer pair, not a verified fact — treat a match as a lead to check, not a conclusion. It also only ever compares a candidate against nodes found by embedding similarity; a contradiction from an unrelated-sounding node would never surface. Comprehensive contradiction detection (not just for the pair the vector search happens to surface) is still an open problem, and nothing here supersedes a node on its own.

provenance is a separate question from trust_state: provenance says where a claim came from, never whether anyone checked it.

nexusmem review <nodeId> --verify
nexusmem review <nodeId> --reject

Records your own verdict on one node, independent of the SLM contradiction checker above (which only ever writes a suggestion, never a verdict). --reject down-weights the node in ranking — same demote-not-delete rule as mark-stale, it stays queryable, just usually loses to better matches — and both verdicts are shown as a [verified]/[rejected] tag on every query result that returns the node afterward. --verify is a label only; it does not boost ranking. Every node starts candidate (untagged) until reviewed, and a re-sync never overwrites a verdict once one is set.

Where it breaks

• Shell history without the hook is unscoped. Scraped history has no directory context, so it is attributed to whichever repository you ran sync from. Bounded to a tail window, and an approximation rather than a guarantee.
• Japanese and Chinese depend on the vector pass. FTS5's unicode61 tokenizer splits on whitespace, so languages without space boundaries get no useful BM25 recall.
• Rebasing strands nodes. Rewritten history leaves nodes for unreachable commits. They describe real events so they are not wrong, but a targeted prune does not exist yet. sync --rebuild forces a clean re-scan.
• Multi-line PowerShell input is read as separate commands. A function typed across several lines at the prompt is not reconstructed.
• Scrape-fallback ids drift if the history file is trimmed from the front between syncs. Installing the hook fixes this.
• Session-summary titles depend on the model following instructions, and a 3B model often does not. The fallback keeps them specific rather than generic, but see the section above for what to expect.
• Changing the embedding model re-embeds everything. Vectors from two models are not comparable and nodes_vec records no per-row provenance, so sync drops the lot and rebuilds rather than ranking across a mixture. It says so when it happens. Nodes are untouched and BM25 keeps working throughout.
• Diff indexing is bounded, and deliberately lossy. A first sync indexes the patches of the most recent 200 commits (later syncs only walk cursor..HEAD); merge commits contribute none, since their patch exists only in a combined format this parser does not read; and binaries, lockfiles and build output are skipped so a dependency bump cannot bury the corpus. All of it is still recorded as a git_commit node. A patch longer than limits.maxBodyChars is truncated, so the tail of a very large change is not indexed. The caps live under sources.diff in config.json.
• Cross-project recall favours breadth. Each repository's hits are fused by rank, so a project whose best match is mediocre still contributes a rank-1 item, and rank 1 is worth the same in every list. Adding a repository that has little to say about your question still pushes a few of its results into the budget. Signal, recency and the budget are what hold that in check; there is no per-project quality weight.
• The project registry is an index, not a source of truth. It can point at a database that has moved or been deleted; those are reported and skipped, never silently pruned, because an unmounted drive is not a deleted project.
• Conversation chunking is unevaluated. Splitting long replies at heading boundaries measurably helped, but it has never been tested systematically.
• A chunked node's sibling count in one result is capped, not tuned. conversation_turn and doc_section both split one reply or file into several nodes; at most 2 of them may appear together in a packed result. Found live: a query for "token" returned 9 of its top 12 hits as different pieces of one heavily-sectioned reply, crowding out the node that actually answered it. The cap of 2 is a judgement call, not a measured optimum, same as the ranking priors' budget above.
• The size of the prior budget is a judgement call, not a measured optimum. Priors are now bounded jointly rather than one at a time, which closed a real 4× hole (see the ranking section), but the 2× budget itself has never been tuned against a labelled relevance set — there isn't one. It is a defensible constant, not a result. What is measured is the direction: on four real queries against this repo's own memory, switching to the joint cap moved the section that answered the question up in three of them (the rationale section for "why BM25 before vector search" went from rank 4 to rank 1) and displaced no query's correct top hit.
• forget is per-repository, not global. Its deny-list lives in the one .nexusmem/memory.db it ran against (plus that repo's stale prior identities, same scope --prune-source already uses). A value that leaked into shell history from several repositories needs forget run once per repo — there is no shared, machine-wide deny-list across every project you have synced.
• A deny-list doesn't survive a clone or restore on its own. .nexusmem/ is gitignored by design, so deny_list never travels with git clone/git push — while git history itself, the thing a fresh sync re-derives from, is fully portable and copied by every clone. A teammate's fresh checkout, a new machine, or a restored backup starts with zero protection: the forgotten value comes right back on the first sync. Confirmed live 2026-08-17, not just a theoretical read of the code. forget --export <path> / forget --import <path> close this: export writes the active entries to a plaintext JSON file you move through a channel you control (never git — the file is exactly as sensitive as the value it holds), and import re-applies them in the new checkout, deleting any copies that already synced back in. It is deliberately manual, not automatic on every sync.

Commands

init, sync, query <text> (add --as-of <date> for a bi-temporal read, see below), status (add --share for a plain-text summary worth pasting somewhere), projects, mcp, forget <value>, stale (add --check-contradictions for a local-SLM content check, see above), mark-stale <nodeId> --supersedes <newNodeId>, review <nodeId> --verify|--reject (record a human verdict on one node, see above), precheck (advisory — warns about staged files with an unresolved past failure or high recent churn; exits 0 unless --strict), hook install|remove|status (the PowerShell exit-code hook), hook git install|remove|status (a git pre-commit hook that runs precheck before each commit), and hook git-post install|remove|status (a git post-commit hook that runs a full sync, including embedding, in the background after each commit — detached, so it never makes git commit itself wait; a burst of commits coalesces into one sync via sync --auto's lock instead of piling up).

There are also eight dry-run previews (scan-git, scan-diff, scan-shell, scan-docs, scan-conversation, scan-session, scan-github, scan-structure) that write nothing and print what ingestion would produce — nodes and their signal scores for the first seven, import-graph edges for scan-structure. That is the intended way to tune scoring against a real repository before committing to a change. Add --json to pipe them somewhere.

Every command takes -C <path> to target another repository. On sync, --conversation and --github opt their (opt-in) sources in for one run without persisting it, --no-embed skips the vector pass, --link-failures builds the failure → fix chains described above, and --rebuild drops the project's nodes and re-ingests from scratch.

sync --prune-source <name> deletes an entire collector source (e.g. shell:pwsh); forget <value> is the finer-grained complement — it deletes every node matching one exact string (or --regex pattern) and writes a standing deny-list entry so the value can never be re-ingested, even by a later sync --rebuild re-reading the append-only shell-hook log or a full transcript scan. Every removal leaves a hash-only tombstone, never the forgotten content itself. Both are dry-run by default; --yes confirms. forget --list shows active entries; forget --export <path> / forget --import <path> carry them to another checkout of the same repo (see the limitation above). See docs/forget-mechanism.md for why this exists.

Recall across projects

query --all-projects searches every repository you have run NexusMem in, not just the current one, and tags each result with the repository it came from:

$ nexusmem query --all-projects "why was the retry budget raised"
scope   2 project(s): NexusMem, uploader

- 2026-08-12 [observed] [uploader] fix: raise the retry budget after the S3 upload timeouts
- 2026-08-12 [observed] [uploader] retry.ts @ 8d0f98b — fix: raise the retry budget after the S3 upload timeouts
  @@ -1 +1 @@
  -export const RETRY_BUDGET = 3;
  +export const RETRY_BUDGET = 5;
- 2026-08-09 [observed] [NexusMem] fix(git): retry a transient failure to spawn git

Bi-temporal reads

Every node carries two clocks: ts, the event's own time ("what happened then"), and created_at, the moment the store actually recorded it ("what did the store hold then") — normally the same question, but not when a sync runs late, a backfill lands weeks after the events it describes, or a teammate's clone catches up all at once. query/search_memory answer the first by default; --as-of <date> switches to the second:

nexusmem query "why does the ranker cap joint priors" --as-of 2026-08-10

Only nodes recorded at or before that instant are considered, even if the events they describe are older still. There is no equivalent write — this is a read-time filter over created_at, not a snapshot or a way to query a value that has since changed, since nodes are write-once (see Staleness & provenance for what changes a node's weight, not its record).

Databases stay per-repository — there is no shared global store, and deleting one repo's .nexusmem/ still removes exactly that repo's memory. What makes the others findable is a plain index at ~/.nexusmem/projects.json, written by init and refreshed by every sync. nexusmem projects shows what is in it, and --prune forgets entries whose database is gone.

Ranking across repositories uses reciprocal rank fusion per project rather than raw BM25, because a BM25 cost is computed against its own corpus and means different things in a 50-node and a 50,000-node database. The trade is stated in Where it breaks.

The MCP search_memory tool takes the same switch as allProjects: true.

On disk

<repo>/.nexusmem/
  .gitignore     '*' — the workspace ignores itself, so init never edits a file it doesn't own
  config.json    validated on read; a corrupt config fails loudly rather than silently
  memory.db      SQLite in WAL mode

~/.nexusmem/
  projects.json      which repositories exist, for cross-project recall; a corrupt one reads as empty
  shell-history.jsonl  the hook's log, if you installed it

NEXUSMEM_HOME overrides the user-scoped directory.

Node ids are content-addressed from sha256(projectId + kind + naturalKey), so running sync twice cannot produce duplicates and ingestion stays correct even if a cursor is lost. Project identity comes from the normalized origin URL when there is one, falling back to the absolute path, so two clones of the same repo share one memory namespace.

Deleting .nexusmem/ loses nothing that sync cannot rebuild.

Status

Ingestion, hybrid retrieval, budgeted packing and the MCP server all work and are covered by 701 tests running on Linux and Windows across Node 22 and 24.

Development

npm install
npm run typecheck
npm test
npm run build

Tests are behavioral rather than snapshot-based, and several are regressions tied to specific observed failures. tests/git-errors.test.ts injects a fake spawn to exercise the Windows process-spawn faults, which cannot be provoked on demand.

Bug fixes, test coverage, and small focused features are welcome — see CONTRIBUTING.md for the workflow, or browse issues labeled good first issue for something scoped and self-contained to start with.

On how this was built

This started as an experiment in whether a local context-memory engine for coding agents was viable, prototyped with Claude Code. The code was written through AI-assisted workflows; the architecture, the design decisions and the specifications were human-directed.

That is worth stating plainly because it should change how you read the code, not whether you trust it. Audits, corrections and PRs are genuinely welcome, and the commit history is deliberately detailed about why things are the way they are, including the times an earlier assumption turned out to be wrong.

License

MIT