deja-vu turns the gigabytes of local conversation logs that Claude Code, Codex, and opencode already write into a searchable memory layer. A single zero-dependency binary indexes existing histories — including months of past sessions — and provides sub-10 ms lexical search across 3.3 GB corpora. Features include an MCP recall tool so agents answer "we fixed this three weeks ago" instead of re-debugging, a Session Start hook that injects relevant context automatically, and credential redaction at index time (API keys, JWTs, private keys, provider tokens).

The tool exports sanitized session digests for colleagues (deja share) and syncs memory between machines via a shared folder or SSH (deja sync), with append-only, idempotent JSONL batches. deja stats summarizes harness usage, top projects, and activity sparklines. The index occupies roughly 2.4 % of the source corpus and updates incrementally.

Nothing leaves your machine: no network code exists in the indexing or search paths. Installation is one command (curl, Go, npm, or Homebrew); deja install --all wires MCP recall and auto-recall hooks into all supported agents. Add a reminder to your CLAUDE.md or AGENTS.md to have agents query memory before debugging.