🔁

Recall — fully-local project memory for Claude Code

Claude Code starts every session cold. Recall keeps a local log of your sessions and condenses it into a resume-ready summary — entirely on your machine. No API key, no external model, nothing sent anywhere. It's built for people running Claude Code locally on a subscription: the only AI in the loop is Claude Code itself; the summarization is done by a classical Python summarizer.

Why Recall

• Free on your subscription. It solves the cold-start problem — no more re-explaining the project each session — without a metered summarizer running up a bill. The summary is a local algorithm, not an LLM call, so persistent memory costs you nothing beyond the subscription you already pay for.
• Saves your usage credits. Two ways: (1) the summary is built locally, so capturing and updating your memory spends zero model tokens; and (2) resuming from a compact context.md (~1–2K tokens) instead of re-explaining the project from scratch each session means far fewer tokens spent per session — stretching your subscription's usage limits (or, on the API, lowering billed credits).
• Nothing leaves your machine. Your transcripts (code, paths, sometimes secrets) are never sent to any API. Most "memory" tools pipe your context to a model endpoint; Recall makes a privacy guarantee they can't. See PRIVACY.md for the full policy.
• Zero-friction. No pip install, no local model to run, no key to configure, works offline. It starts working the moment the plugin loads.

Two files, written into your project under .recall/:

• history.md — the log. Append-only. Every session is captured here as it happens (your prompts, Claude's replies, the files touched and commands run).
• context.md — the summary. Overwritten by the local summarizer — the condensed "where are we right now" you load into the next session: goal, summary, next steps / open threads, files touched, and where you left off.

"Doesn't Claude Code already have memory?"

It does — and Recall is complementary, not a replacement. The built-in options solve different problems:

• CLAUDE.md (and the # shortcut) is hand-written memory: rules and notes you curate, loaded as instructions Claude follows. Great for "how I want you to work," but it's manual upkeep and it doesn't record what actually happened in a session.
• --continue / --resume replays a prior conversation — full fidelity, but it reloads the whole transcript (token-heavy) and is tied to your local session history on one machine, not a portable, readable digest.
• Context compaction condenses a conversation within a session; it isn't a durable record you reopen days later.

Recall fills the gap between these: an automatic, deterministic record of what each session did, condensed into a compact resume point.

In short: CLAUDE.md is how I want you to work; Recall is here's what we did last time and where we stopped — produced offline, with no model tokens spent.

How it works

There is no LLM call anywhere — the summary is produced by TF-IDF + TextRank (extractive summarization) running locally.

The summarizer

scripts/summarizer.py ranks the most central sentences of your session:

1. TF-IDF sentence vectors
2. a cosine-similarity graph between sentences
3. TextRank — PageRank power iteration over that graph — to score sentences
4. the top N are kept in original order

context.md wraps that summary with deterministic facts pulled straight from the transcript and git: the goal (your first ask), files touched, commands run, where you left off, and git diff --stat.

No installs required. The whole TF-IDF + TextRank implementation is vendored in summarizer.py. If numpy happens to be importable it's used to vectorize the math (faster on big sessions); if not, an identical pure-Python TextRank runs instead. Same algorithm, same result — numpy is an optional accelerator, never a requirement. The save output tells you which path ran.

Commands

• /recall:save — run the local summarizer → (over)write context.md.
• /recall:show — print context.md.
• /recall:log — tail history.md.

Configuration — recall.config.json

Drop this in your project root to override defaults:

Pause logging for a project without editing config: create .recall/.capture-paused. Delete it to resume.

Privacy & security

Recall makes no network calls, uses no API key, and loads no third-party model. The summarizer is local Python; the hooks are stdlib-only (numpy is an optional accelerator). It reads your session transcript and writes only under output_dir. Concretely:

• No credentials, ever. The plugin has zero references to API keys, auth, ANTHROPIC_*, or HTTP. If claude itself shows "Invalid API key", that's the CLI's own auth — usually a stale ANTHROPIC_API_KEY env var shadowing your subscription login. unset ANTHROPIC_API_KEY (or run env -u ANTHROPIC_API_KEY claude …). It has nothing to do with Recall.
• Redaction. A best-effort pass strips common secret shapes (API keys, tokens, .env assignments, PEM keys) before writing, since context.md / history.md may be committed. Best-effort, not a guarantee — review before committing.
• Hardened git. git diff/log are run with core.fsmonitor, diff.external, hooks, and the pager disabled, so an untrusted cloned repo can't use its own git config to execute code when Recall reads ground-truth. Set include_git: false to skip git entirely.
• Confined writes. output_dir is forced to stay inside the project; a project-shipped config can't redirect writes to an absolute path or ../...
• Scoped transcript. Recall only reads the transcript for the current project (matched by cwd); it never falls back to another project's sessions.
• Trust boundary for shared memory. context.md is injected into the model at session start. If you commit .recall/ as shared team memory, treat it like any other shared input: a teammate (or a bad actor with repo write access) could craft a context.md to attempt prompt-injection. SessionStart fences the content and labels it untrusted data, and Claude asks before relying on it — but if you don't fully trust who can write the repo, keep .recall/ git-ignored (the default).

Committing .recall/

Both are fine. Commit it for shared team memory, or git-ignore it for personal memory (.gitignore ships ignoring it by default — flip the comment to commit).

Install

From the marketplace (this repo is its own marketplace):

/plugin marketplace add raiyanyahya/recall
/plugin install recall@recall

Local dev (no install step):

claude --plugin-dir /path/to/recall

No pip install — the summarizer is vendored and stdlib-only (numpy used as an optional accelerator if present). Work a session, run /recall:save, and open a fresh session — Recall greets you with where you left off.

Development

python -m venv .venv && . .venv/bin/activate
pip install pytest ruff bandit numpy   # numpy optional

ruff check scripts tests benchmarks    # lint
bandit -c pyproject.toml -r scripts    # security static analysis
pytest                                 # run the suite (also test without numpy)
python benchmarks/bench.py             # perf + quality numbers (human-readable)
python benchmarks/bench.py --check     # assert quality invariants (the CI gate)
claude plugin validate .               # official manifest validation

benchmarks/bench.py is a stdlib-only harness: alongside latency/throughput it scores the summarizer's salient-sentence selection against lead/tail/random baselines on a labeled fixture set and checks the numpy and pure-Python cores select the same sentences. --check gates those quality invariants (it never gates wall-clock timings). Redaction quality is covered by the unit suite (tests/test_redact.py), so no secret-shaped fixtures live in the benchmark.

CI (.github/workflows/) runs lint + Bandit, the test suite across Python 3.9–3.13 with and without numpy (both summarizer paths), the benchmark quality gate (both paths), CodeQL, secret scanning, and manifest JSON validation on every push and PR. See CONTRIBUTING.md and SECURITY.md.

Layout

recall/
├── .claude-plugin/plugin.json   # manifest
├── hooks/hooks.json             # SessionStart (ask/resume) · Stop+SessionEnd (capture)
├── commands/                    # /recall:save · show · log
├── scripts/
│   ├── summarizer.py            # vendored TF-IDF + TextRank (numpy optional)
│   ├── make_context.py          # build/overwrite context.md
│   ├── capture.py               # append session activity to history.md
│   ├── session_start.py         # surface context + ask the start questions
│   ├── parse_transcript.py      # transcript → events + renderers
│   └── config.py · common.py · redact.py
├── tests/                       # pytest suite (summarizer, capture, security, …)
├── benchmarks/bench.py          # perf + quality harness (CI quality gate)
├── .github/                     # CI, CodeQL, secret scan, dependabot
├── recall.config.json        # config template / defaults
├── pyproject.toml               # ruff / pytest / bandit config (no runtime deps)
├── LICENSE · SECURITY.md · CONTRIBUTING.md
└── .gitignore

Contributing & issues

Bugs and ideas are welcome — open an issue (bug-report and feature templates provided) or a pull request. See CONTRIBUTING.md before submitting, and report security vulnerabilities privately per SECURITY.md rather than in a public issue.