Recall: a fully local memory plugin that fixes Claude Code's cold start

Recall is a Claude Code plugin that tackles the cold-start problem—having to re-explain your project at the top of every session. It maintains two files under a project’s .recall/ directory: an append-only history.md log capturing prompts, replies, files touched, and commands run, plus a context.md digest summarizing the current goal, recent work, next steps, and where you left off. A fresh session loads the compact context.md (roughly 1–2K tokens) instead of replaying a full transcript.

The distinguishing claim is that no AI is involved in the memory itself. Rather than piping context to a model endpoint like many ‘memory’ tools, Recall builds its summaries with a vendored, classical extractive algorithm—TF-IDF sentence vectors fed into a cosine-similarity graph and ranked by TextRank (PageRank iteration). It’s stdlib-only Python, with numpy as an optional speed-up, so there are no installs, API keys, or network calls. That makes summarization free against an existing subscription and saves model tokens twice over: capturing memory costs nothing, and resuming from a short digest trims per-session token spend.

Security gets unusually thorough treatment for a Show HN. Recall does best-effort redaction of secret-shaped strings before writing, runs git with external diff tools, hooks, and config-driven execution disabled to neutralize untrusted cloned repos, confines all writes inside the project directory, and scopes transcript reads to the current project. Because context.md is injected into the model at session start, the author flags shared team memory as a prompt-injection surface—SessionStart fences it as untrusted data and Claude asks before relying on it, with the safer default being to git-ignore .recall/. It’s positioned as complementary to CLAUDE.md and —resume rather than a replacement.