Claude Code forgets. This plugin remembers—offline and for free.

What it does

Recall is a Claude Code plugin that appends every session to a local history.md log, then condenses it into a compact context.md summary using classical extractive summarization. When you start a new session, it surfaces that summary so Claude knows where you left off without you re-explaining the project. It runs entirely on your machine with no API calls, no external models, and no keys to manage.

The interesting bit

Instead of paying an LLM to summarize your transcripts, Recall uses a vendored TF-IDF + TextRank implementation—pure Python with an optional numpy accelerator—to pick the most salient sentences. That keeps your session memory inside your subscription’s usage limits instead of burning extra tokens on a meta-LLM call.

Key highlights

• Entirely offline: no network calls, no API keys, no third-party models; transcripts never leave the machine.
• Zero token cost for memory: the summarizer is a classical algorithm, not an LLM, so capturing and updating context costs nothing beyond your existing Claude Code subscription.
• Complements built-in memory: unlike CLAUDE.md (hand-written rules) or --continue (reloading full transcripts), Recall produces an automatic, compact digest of what actually happened.
• Privacy-minded: best-effort redaction strips secret shapes before writing, git commands are hardened against untrusted repo configs, and writes are confined to the project directory.
• Portable and diffable: history.md and context.md live under .recall/ as plain Markdown you can commit for shared team memory or keep git-ignored for personal use.

Caveats

• The summarizer is extractive, not abstractive: it ranks and pulls existing sentences rather than synthesizing new meaning, so the quality depends on what you literally said.
• Redaction is best-effort, not a guarantee; the README explicitly warns you to review before committing if secrets were discussed.
• Shared team memory carries a prompt-injection risk: if you commit .recall/ to the repo, anyone with write access could craft a malicious context.md, though the plugin fences it as untrusted data.

Verdict

Worth a look if you use Claude Code daily and are tired of the cold-start ritual; skip it if --continue or terse CLAUDE.md notes already cover your workflow.