Deja Vu (the deja command) gives coding agents a shared memory built from the session transcripts they already write to disk. It reads the stored sessions of 41 harnesses, including Claude Code, Codex CLI, Cursor, Gemini CLI, opencode, Goose, Hermes, and OpenClaw, into one local index, so months of history from before it was installed are searchable on day one, and a fix found in one agent comes back in another.
The index is a local inverted index over the harnesses' JSONL and SQLite stores. It needs no model, no embeddings, and no server, and repeat runs re-read only the session files that changed. Credentials (provider keys, tokens, JWTs, PEM blocks, passwords) are stripped as the index is built, and the network is touched only for updates, machine-to-machine sync over SSH, and optional embedding calls. Besides messages, it indexes the files each turn opened, the commands with their exit status, and the exact spans an edit replaced.
Agents reach it through an MCP server that exposes a single deja tool with modes for recall, context digests, file blame, past fixes for an error, how a command was run, project orientation, durable notes, and handoff between agents. With auto-recall, relevant history arrives at session start, before a file is edited, and after a command fails. A terminal search screen, a CLI, and a skill cover the same search for people and for agents that load skills.
Optional semantic recall points the index at a local Ollama or LM Studio runtime, OpenAI, or another OpenAI-compatible embeddings endpoint, so rephrased queries still match. Published results include 97.2% recall at 5 and 88.1% hit at 1 on LongMemEval-S, 70.5% hit at 1 on LoCoMo retrieval, and median in-process lookups of 0.7 to 0.8 ms on a store of 2,419 sessions.
Features
- Cross-agent memory: one index over the session history of 41 harnesses, so a fix found in Codex comes back in Claude Code, Cursor, or opencode
- Day-one history: sessions recorded before installation are searchable immediately
- MCP tool: one deja tool with recall, context, blame, fix, how, orient, remember, and handoff modes
- Auto-recall: relevant history arrives at session start, before a file is edited, and after a command fails
- Work indexing: files opened, commands with exit status, and the spans each edit replaced, not only the conversation
- Decision state: promoted notes mark a decision accepted or rejected, and later hits report a reverted approach and why it was dropped
- Staleness signal: a hit reports when files the session touched have changed since
- Terminal search: a full-screen search across every agent's sessions, with filters by agent, date, and project, and resume or continue in another agent
- Credential redaction: keys, tokens, JWTs, PEM blocks, and passwords are stripped from the index, and a secrets command finds them in the source transcripts
- Sync and handoff: moves memory between machines over SSH with no cloud, and hands a session's goal and conclusions to another agent
- Optional semantic recall: embeddings from Ollama, LM Studio, OpenAI, or an OpenAI-compatible endpoint, with lexical search unchanged without them
- Harness packages: plugins for opencode, Codex CLI, Grok Build, OpenClaw, pi, Hermes, Kimi Code, Zed, and DeepSeek Harness, plus a search skill
