memory-pulse

Site: pulse.strategic-innovations.ai, where the guard runs in the browser. Compare with Mem0, Zep, Letta and CLAUDE.md.

Corrections that outlive the session.

Your agent acknowledged the correction, then wrote the old value back next session. memory-pulse records it once, puts it at the top of every session, and blocks the edit that reintroduces it, quoting the line that retired it. For Claude Code, Codex, Cursor, any MCP client and GitHub Actions. The record is a file in your own repository.

Most memory tools stop at showing the correction to the model. The guard is the part that acts on it.

All four tool definitions come to ~3.3 KB, about 900 tokens, and a test in this repo fails the build if they grow past that. Independent measurements put a typical five to ten server MCP setup at 50-67k tokens of tool definitions before your first prompt, roughly a third of a 200k context window.

What it does

A session ends and everything it learned goes with it. memory-pulse keeps a ledger of cause → effect events in a local file, and gives the agent four tools over it:

Two design decisions do the heavy lifting.

Corrections come first, always. An event recorded with kind: "correction" outranks everything at every brief size and never decays. The failure this prevents is your agent confidently quoting the benchmark number you withdrew three sessions ago.

A weak answer returns nothing. When recall cannot clear its confidence floor it returns an empty result, and the brief says so.

Install

One repo, two plugin hosts, any MCP client. Pick the row for your agent.

Claude Code, as a plugin. Skill, MCP tools and the two hooks, from this repo:

/plugin marketplace add t-crew/memory-pulse
/plugin install memory-pulse@memory-pulse

claude plugin details memory-pulse shows what you pay. About 120 tokens are always on, which is the skill's description. The four tool schemas resolve at runtime and the hooks are free. The hooks run the plugin's own server.mjs, so what enforces your corrections is the version you installed.

Codex CLI, as a plugin. Same files, read from .codex-plugin/plugin.json:

codex plugin marketplace add t-crew/memory-pulse
codex plugin add memory-pulse@memory-pulse

Then, inside Codex, run /hooks and trust the two memory-pulse entries. Codex runs no hook it has not shown you, and installing a plugin does not trust its hooks. That is Codex's rule and a good one.

Any MCP client, in one line. Claude Code and Codex are shown, and Cursor and the rest take the same stdio command:

claude mcp add memory-pulse -- npx -y memory-pulse
codex  mcp add memory-pulse -- npx -y memory-pulse

Automatic re-entry without the plugin. A SessionStart hook runs the brief before your first prompt and a PreToolUse guard checks edits. Both are idempotent, both merge into the settings file without clobbering it, and both stay silent in a project that has no ledger.

npx memory-pulse install-hook            # Claude Code: ~/.claude/settings.json
npx memory-pulse install-hook --codex    # Codex:       ~/.codex/hooks.json (then /hooks to trust)

Add --project to either and the hooks are written into the repo, at .claude/settings.json or .codex/hooks.json. Commit that and every clone is re-entered and guarded with nothing for anyone to install.

A withdrawn number recorded with kind: "correction" outranks the history that contained it, at every brief size and in every session.

Corrections are enforced, not only surfaced. Showing an agent a correction is measurably not enough, because agents re-violate corrections they were just shown. The PreToolUse guard sees every Edit and Write in Claude Code and every apply_patch in Codex, where one patch may touch several files and each is checked under its own path. An edit that writes back a withdrawn value is blocked, and the agent is told which ledger line retired it and when. An edit that names the replacement beside the old value passes, because "was $49, now $29" is a comparison. Only a bare reintroduction is blocked. A shell heredoc is not an edit tool and is not guarded, so check --ci on the pull request is the layer that catches that case. Record corrections with the exact terms:

remember({ cause: "pricing-shipped", effect: "price-corrected", kind: "correction",
           note: "measured willingness to pay is $29", withdrawn: ["$49"], replacement: ["$29"] })

Commands

npx memory-pulse brief          # the re-entry brief (what the SessionStart hook prints)
npx memory-pulse brief --budget 1500   # size it to the tokens you can spare; the richest tier that fits, corrections first and whole
npx memory-pulse guard          # PreToolUse hook: blocks edits that reintroduce withdrawn terms
                                # (a later correction can `supersedes: [t]` an earlier one — only the latest binds)
npx memory-pulse check --ci     # Memory CI: one of three verdicts for a change, from files you own
npx memory-pulse verify         # row chain + last engine seal; exit 2 if either fails
npx memory-pulse brief --offline  # local render when the engine is unreachable
npx memory-pulse install-hook --ambient  # also record prompts shaped like corrections
npx memory-pulse lint [--ci]    # dry run: do CLAUDE.md / AGENTS.md / .claude/rules still state a value the ledger retired?
npx memory-pulse guard allow "<term>" --path <prefix> "<reason>"   # record a false block as an override
npx memory-pulse report         # correction re-violation scoreboard, computed locally
npx memory-pulse bench          # instant measured metrics on YOUR ledger
npx memory-pulse stats          # your telemetry capsule, signature verified by the engine
npx memory-pulse badge          # README badge markdown from your own signed numbers
npx memory-pulse install-hook   # installs both hooks (idempotent); --codex targets Codex; --project commits them to the repo

The plugin also ships a skill at skills/memory-pulse/SKILL.md that teaches the agent when to pulse, how to record corrections with withdrawn terms, and how to respect the guard.

What the brief tells you before it tells you anything

Every brief opens with one line of provenance, so a session can tell whether its memory loaded whole, truncated, or not at all:

memory-pulse: loaded 852 events from .memory-pulse/events.jsonl · sha256 1a2b3c4d5e6f · 2 binding corrections (10 withdrawn terms) · 1 superseded · ⚠ 1 malformed line skipped: 544 · memory key resumed (+3 new) · tier brief, 5,153 chars

Every CORRECTIONS line cites its ledger record as … -> effect (t12) — note, so a correction is evidence the agent can point at. recall and the guard name the same t.

Lint: the rules a session loads, checked against the ledger

Governance files drift. A CLAUDE.md written in June still says the price is $49 after the ledger retired it in August, and every new session loads the stale rule with full confidence. lint runs the guard's check over the files a session will read and gives each of them one of the three verdicts. It covers CLAUDE.md, AGENTS.md, .claude/rules/, .cursorrules, .cursor/rules/, .github/copilot-instructions.md, .codex/AGENTS.md, and any paths you pass:

$ npx memory-pulse lint
memory-pulse: loaded 2 events from .memory-pulse/events.jsonl · sha256 8e401a39f323 · 1 binding correction (1 withdrawn term)
  BLOCKED     CLAUDE.md
             • "$49" was withdrawn at ledger t2: price-49-launched -> price-corrected-to-29 — use $29
  verified    AGENTS.md
  no evidence .claude/rules/style.md
lint: 3 file(s) — 1 blocked, 1 verified, 1 no evidence — a rule your ledger retired is still being loaded into sessions

Exit 2 on any blocked file. Under --ci it also exits 1 when it found nothing to check, so an empty run is never reported as a pass. --json is there for machines. It also names the corrections that carry no withdrawn terms, which surface in the brief but which nothing can enforce.

Tamper evidence: the ledger cannot be edited quietly

Since 0.3.1 three mechanisms hold, each with its own job:

• Row chain. Every row remember writes carries prev, the previous chained row's hash, and hash, the SHA-256 of its own canonical JSON. Rows that existed before the chain are never rewritten, and the first chained row seals them with a digest. An edit in place, a removed row, a reordered row or an unchained row after the chain started all fail verifyChain(). A failed chain blocks every check, guard and lint verdict, because a memory whose own history is in question cannot vouch for anything.
• Set head. The engine also commits to the ledger as an order-free fold, a multiplicative group mod a 3072-bit prime, which is the MuHash construction Bitcoin Core uses for its UTXO set. Shards from several agents fold to the same head in any order, and removing a row is the group inverse, so the state stays exact while the history stays append-only. A literal XOR fold was measured forgeable, since a linear system hides an edit in 10 ms at 300 rows, and it is not used here.
• Seal. Every read call returns a seal signed by the engine, carrying the row count, the watermark, the set head over every row's full content, and the chain head. The client keeps it in seal.rain beside the ledger and presents it on the next call. Locally, rows up to the sealed watermark must fold to the sealed head before the ledger is trusted. At the engine the signature and the fold are re-checked, and an edit below the watermark is reported as drift and blocks. A process with write access can rewrite the file and even the chain. It cannot produce the engine's signature, and it cannot make edited rows fold to the sealed head.

Nothing is stored server-side for any of this. The seal travels in the payload, the same way the telemetry capsule and the memory key do.

Survives compaction, works offline, captures corrections, speaks Python

Four things added on 2026-09-03, each deterministic (no model in the loop):

• Compaction handoff. install-hook adds a PreCompact hook. Before Claude Code compacts, memory-pulse handoff reads the transcript and records what the session was doing as facts: the last asks, the files edited, the last error, and the assistant's last state. The next session start prints it first, online or offline. An instruction-like message is dropped from the note and never recorded.

• Offline brief. When the engine is unreachable, whether air-gapped, on a dead network or during an outage, the session no longer starts empty. brief prints a local render carrying every binding correction with its withdrawn and replacement terms, the last handoff and the recent rows, labelled as a local render with no ranking applied. brief --offline forces it. Guard, check, lint and verify never needed the network.

• Ambient correction capture, opt-in. install-hook --ambient adds a UserPromptSubmit hook. A prompt shaped like a correction, such as the price is $29 not $49, change 0.3.1 to 0.3.2 or 500 events -> 924 events, is recorded as a correction carrying both terms, so the guard enforces it from the next edit on. A prompt that does not yield both terms is left alone. Silent unless --verbose.

• Python client. python/memory_pulse.py is a single stdlib-only file with the same ledger format, the same hash chain and the same guard rule. A LangChain or CrewAI agent and a Claude Code session can share one ledger and verify each other's rows. The test suite writes rows from Python and verifies them in Node, and back.

• Decode-time guard for local models. python/span_guard.py applies the same rule one layer down. The token that would complete a withdrawn value is masked while the model decodes, so the value cannot be generated, and the replacement's next token is offered in its place. It reads the ledger under the rules check uses, so a superseded correction is not enforced and a broken chain refuses to build. Stdlib only, and its suite runs without a model. Measured with mlx-lm on TinyLlama-1.1B and Qwen2.5-1.5B, the same prompt whose plain decode wrote the withdrawn price wrote the corrected one under the guard, and the sentence around it stayed intact. It blocks the spellings it was given, including the written variants you record, and a paraphrase nobody enumerated gets through.

  server.ts

  Copy

  from memory_pulse import Ledger
  from span_guard import SpanGuard
  guard = SpanGuard.from_ledger(tok, Ledger(".memory-pulse/events.jsonl"))
  # mlx-lm: generate_step(..., logits_processors=[guard.logits_processor(mx, np)])
  

Agent mode: a persistent agent identity that grows (opt-in)

The default is deliberate. Memory lives in the repo and capture is explicit. Agent mode covers the other thing people ask for, which is an agent that is the same agent tomorrow, in every project and every tool, and that grows.

npx memory-pulse mode agent
npx memory-pulse identity "Blue, research agent for Travis; innovate, don't debate"

What that turns on:

• An agent ledger at ~/.memory-pulse/agent/events.jsonl, belonging to the agent and shared by every project and every tool that speaks MCP. Same format, same hash chain, same seal.
• A self block, first in every brief, online or offline. It carries who the agent is as a pinned line, the standing rules and preferences it has learned, the lessons it carries, and a fingerprint made of the chain head and the engine's seal, so the agent can state which memory it is running on and show it was not swapped or edited overnight.
• Growth, after every turn. A Stop hook deterministically records a stated decision to the project ledger, a user preference or stated lesson to the agent ledger, and any correction shaped like one. It is capped at four rows a turn, tagged ambient, never pinned, and never taken from instruction-like text. Identity itself is only ever set by you or superseded by a correction.
• Corrections that follow the agent. A correction on the agent ledger blocks the same edit in any project.

npx memory-pulse mode deliberate turns the hooks off again and leaves the ledgers in place. remember takes scope: "agent" from any tool.

What runs where (the privacy contract)

• Your ledger is a local file at .memory-pulse/events.jsonl in your project. You can commit it, grep it or delete it.
• remember writes to it directly and works offline.
• Read operations send the ledger's events to the hosted engine over TLS, which computes the answer and forgets the request. The service keeps no database of your memory. State arrives in the request and leaves in the response.
• Telemetry is a signed capsule beside your ledger at .memory-pulse/telemetry.rain. The engine advances it on each read call and hands it back without ever storing it. stats verifies the signature and badge turns it into a README badge. Delete the file and it restarts.
• State persistence, no database. After a read the engine hands back a signed memory key (.memory-pulse/memory.rain, git-ignored). The next read presents it and the engine resumes from it, ingesting only the events recorded since. The answer is byte-identical to a full rebuild. Any mismatch, whether edited history, a stepped ledger size or a bad signature, falls back to a rebuild and says why. Lose the file and you lose nothing but one rebuild. MEMORY_PULSE_MEMORY_KEY=off disables it.
• Memory integrity. A note that reads like an instruction, such as "ignore previous instructions", "run this command" or a fake system tag, is refused by remember. One already sitting in a ledger is quarantined at read time and reported, so memory never reaches your agent's context as an order. The signed capsule also raises a drift alert when a ledger loses corrections, shrinks, or its usage shape jumps, and the brief footer shows it. Both checks are deterministic lists you can read.
• This client is the entire client: one file, zero dependencies, readable in one sitting.

Pricing

• Free. Ledgers up to 500 events, 200 reads a day. No account, no key.
• Pro, $19/mo. Ledgers to 20,000 events, unlimited reads. One environment variable, MEMORY_PULSE_KEY.

Local writes are free on either tier.

Measured, on our own ledger

Measured on the ledger of the project that builds memory-pulse, a 767-event file of 1.08 MB, measured 2026-09-01:

• A cross-referencing question answered through execute returned 124 chars against the 1,080,983-char full dump. The intermediates never entered context.
• Re-entry briefs at the smallest tier run ~99% smaller than reading the ledger in.
• On our recall benchmark of 351 distinct causes, the noise-floor gate returned zero wrong top answers. When it could not clear the floor it returned nothing.

The ratios depend on ledger size, and a ten-event ledger has nothing to compress. The methodology lives in the engine's benchmark suite.

FAQ

Why is the engine hosted? The ranking engine is the part that took the research. We chose a local ledger, a thin auditable client and a hosted engine over shipping a weaker local ranker. If a remote engine is a dealbreaker, MEMORY_PULSE_API points the client anywhere.

What about team memory? Commit .memory-pulse/ to the repo and your teammates' agents pulse the same ledger. Shared hosted ledgers are on the roadmap.

License? Client: MIT. Engine: proprietary, hosted.

MIT © Travis Crew

Memory CI: three verdicts, including an explicit empty one

memory-pulse check gives a change one of three verdicts, computed locally from your ledger and your declared invariants:

• blocked. The text writes back a value a correction withdrew, and the verdict names the ledger line that retired it and what to use instead. Also fires when the text trips a declared invariant. Exit 2.
• verified. Recorded events bear on the text and none is contradicted. Exit 0.
• no evidence. The ledger has nothing to say, reported as exactly that and never as a pass. Exit 1 under --ci. The hook stays silent on this verdict, because a hook that comments on every edit livelocks the agent, and CI is where it is loud.

npx memory-pulse check --ci --diff origin/main      # added lines of the branch
npx memory-pulse check --ci --file docs/pricing.md
echo "price is $49" | npx memory-pulse check --ci
npx memory-pulse check --receipt --text "…"        # engine-signed receipt, keyless verify at /v1/verify

Invariants are declared, never inferred. They live one per line in .memory-pulse/invariants.jsonl, shaped like {"id":"receipt-wording","statement":"say tamper-evident", "patterns":["/\\bproof\\b/i"],"paths":["site/"],"severity":"block"}. A pattern written /…/flags is a regular expression and anything else is a verbatim substring. paths scopes the rule to path prefixes, so a rule about public wording does not fire on a proofs file. severity: "warn" reports without blocking.

Overrides are the false-block signal. guard allow "$49" --path docs/history "historical table" records an override event scoped to that path prefix. The hit passes there and nowhere else, and both report and the signed capsule count it. The guard never guesses, and only explicit withdrawn terms and declared invariants can block. Measured on our own 852-event ledger, with the bench in the engine repo: precision 1.0, zero false blocks over 871 negatives including 694 real notes, and p95 of 1.7 ms at 1k events.

PR status check with uses: t-crew/memory-pulse@v0

The same three verdicts as a GitHub check on every pull request, against the ledger and invariants committed in your repository:

# .github/workflows/memory-ci.yml
on: pull_request
permissions: { contents: read, checks: write, pull-requests: write }
jobs:
  memory-ci:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
        with: { fetch-depth: 0 }
      - uses: actions/setup-node@v4
        with: { node-version: 22 }
      - uses: t-crew/memory-pulse@v0

It checks the added lines of the pull request. blocked fails the check, and the sticky comment cites the ledger line that retired each value. verified succeeds with "N memories checked, M corrections enforced". no_evidence is a neutral conclusion with an explanation, so an empty evidence set is never reported as a pass. The comment is one per pull request and updated in place. Nothing leaves the runner unless you pass api-key for signed receipts. This repository runs it on itself, which is why .memory-pulse/events.jsonl is committed, and the first pull request it blocked is the demo.

Releasing

npm version patch && git push --follow-tags

The tag triggers .github/workflows/release.yml, which runs the tests and publishes to npm and the MCP registry using GitHub's OIDC identity. There are no tokens in the repo and no one-time passwords. npm attaches provenance automatically, so anyone can verify the package was built from this repo.

One version has four readers: package.json for npm, server.json for the MCP registry, .claude-plugin/plugin.json for Claude Code and .codex-plugin/plugin.json for Codex. All four must agree. npm version bumps only the first, test/manifests.test.js fails until the other three follow, and the release workflow refuses to publish while they disagree.