Recall — offline session memory for Claude Code

A Claude Code plugin giving durable, fully offline project memory (raiyanyahya/recall; Python; MIT; ~387 stars). It targets the cold-start problem — Claude Code beginning every session with no context — and is the most minimal pole the spoke has paged of agent-memory. Shared via Telegram, ingested 2026-06-23.

How it works

• Delivery: lifecycle hooks, not MCP. SessionStart/Stop hooks capture activity; slash commands (/recall:save, /recall:show, /recall:log) drive it. No MCP server, no separate CLI, no database.
• Storage: plain markdown files. Two files in a .recall/ dir — history.md (append-only session log) and context.md (a regenerated summary). Git-ignored by default but committable for team memory.
• Summarization: deterministic, no LLM. A vendored local TF-IDF + TextRank extractive summarizer (optional NumPy speedup) produces a ~1–2K-token digest — no model call, no embeddings.
• Offline guarantee. “Nothing leaves your machine”; best-effort secret redaction before writing.

Why it matters here — the third pole of agent-memory

agent-memory already had two market poles: a full branch/merge DB (seekdb) and a thin pgvector MCP store (memory-vault). Recall extends the spectrum to its minimal, local-first end on every axis:

• Storage: DB → Postgres+pgvector → plain markdown files.
• Delivery: queried as a DB → handed over MCP → fired from lifecycle hooks.
• Index/summarize: vector + full-text → extractive TF-IDF/TextRank, no embeddings, no LLM (so the digest is deterministic and free, where memory-vault solves the same Claude Code auto-compaction amnesia with a vector store reached as a tool).

The trade is explicit: recall gives up semantic retrieval for zero dependencies, zero cost, and total locality. Its choice to store memory as plain markdown you own (and can commit) is the markdown-you-own ethos applied to session memory — a direct cross-wiki rhyme with research-wiki’s gbrain/llm-wiki “git repo of markdown as system of record.” (Reflexively, the SessionStart-hook-regenerates-context.md loop is the same shape as this hub’s own MEMORY.md.)

Caveats

Primary repo (T2): claims are self-reported; small/young (v0.3.5, ~387 stars). Extractive summaries are deterministic but lossy in a different way than LLM compaction — they pick sentences, they don’t synthesize.

Related

agent-memory · memory-vault · seekdb · durable-agents · anthropic · agent-skills · gbrain