Recall

Every Claude Code session starts from zero. You open the terminal, and Claude has no idea what you were doing yesterday, or twenty minutes ago if you closed the window. Recall, a new open-source plugin, tackles that cold-start problem with a trick that's almost stubbornly old-fashioned: instead of asking an LLM to summarize your work, it uses classical algorithms — TF-IDF and TextRank, the same math that powered search-engine snippets a decade ago — to condense your session into a short markdown file, entirely on your machine.

The mechanics are simple. Recall keeps two files inside a .recall folder: history.md, an append-only log of every prompt, reply, file touched, and command run, and context.md, a compact digest rewritten by the local summarizer that captures the goal, the summary, next steps, and where you left off. When a new session starts, Claude reads that digest — usually 1 to 2 thousand tokens — instead of you re-explaining the project from scratch or Claude replaying a full transcript via --resume. No API key touches this process. No text leaves the laptop.

That local-only design isn't just a privacy flex, though it is that too — the maker points out that most memory tools quietly pipe your context to a model endpoint, something Recall structurally cannot do. It's also a cost argument. Because the summarization runs as plain Python rather than a metered LLM call, capturing and updating memory burns zero extra tokens, and resuming from a short digest instead of a bloated transcript stretches whatever usage limits your subscription gives you. For people running Claude Code on a flat-rate plan rather than pay-per-token API access, that's real money left in your pocket.

What's notable is how deliberately Recall avoids overlapping with Claude's existing memory tools rather than trying to replace them. CLAUDE.md remains your hand-curated instructions file. --resume still replays full conversations when you need that fidelity. Recall just fills the specific gap neither covers: an automatic, zero-effort record of what actually happened, redacted for obvious secrets, hardened against malicious git configs in untrusted clones, and fenced off as untrusted data so Claude doesn't blindly trust a context.md a bad actor might have tampered with in a shared repo. It even extends past Claude Code — an opencode installer lets two different coding agents share the same history.md and context.md on one repo, which is a small but genuinely useful bit of interoperability nobody asked for but everyone building multi-agent workflows will eventually want.

The project ships with a benchmarking harness that scores its own sentence-selection against lead, tail, and random baselines, and CI runs the whole suite with and without numpy to make sure the pure-Python fallback produces identical results. That's an unusual level of rigor for what is, on paper, a memory-jotting plugin — but it suggests the author is serious about this staying boring, deterministic, and dependency-free, which is exactly what you want from something reading your codebase every session.