seek://map/deja-vu
deja-vu — Shared Memory for Coding Agents
#memory #cross-agent #search #mcp
TL;DR
Per its README, deja-vu indexes the session files your coding agents already wrote — Claude Code, Codex, Cursor and more — into a local inverted index any of them can search, including history from before installation; DeepSeek Harness gets its own npm package, dsh-deja, via dsh plugin --profile web add dsh-deja. No model calls, no embeddings, credentials redacted as the index is built.
Per its README's own framing, memory tools start empty and record forward — deja-vu starts full: it is a local search index (vshulcz's Go binary, plus an MCP server and a CLI) over the session files Claude Code, Codex, Cursor, VS Code Copilot Chat and more already wrote to disk, sessions from before installation included. DeepSeek Harness is a supported row in its matrix, with its own dsh-deja npm package, a dedicated guide page, and a needs column that asks for zstd beside marked limits: no session resume, paste-only handoff. Recall is a lexical lookup that never waits on a model, credentials are stripped as the index is built, and the benchmark harnesses ship in the repo so the published numbers can be re-run. That cross-agent reading list is what earned the map spot.
Facts
- license
- MIT
- note
- DSH 矩阵行标注：需要 zstd；resume 不可用、handoff 仅粘贴。README 自述密钥脱敏是模式匹配，不认识的形态可能漏过。
Key points
- Per its README, the index is built from session files already on disk — sessions from before installation included — so a decision from months back is searchable in whichever agent asks
- DeepSeek Harness has its own row in the support matrix and its own npm package: install dsh-deja with dsh plugin --profile web add dsh-deja; the matrix marks MCP recall, auto-recall, skill and command as working, resume as unavailable and handoff as paste-only, with zstd required
- Local and model-free: recall is a lexical lookup against a local inverted index in ~/.cache/deja — per its README, nothing leaves the machine unless you ask
- Credentials are stripped while the index is built; the README's own caveat says pattern matching is not secret detection, and a shape it does not know can pass through
- Per its README's benchmark page: 85.3% hit@1 on LongMemEval-S, 69.6% on LoCoMo, and ~0.4 ms median in-process lookup
- Beyond recall, the CLI answers workflow questions: blame shows which sessions touched a file, fix shows what ran after this error before, friction lists errors that keep recurring, and restore hands back a span an agent replaced
FAQ
How does it attach to DeepSeek Harness?
Per its README, either path is enough on its own: deja install --auto wires it like every other harness, or the dsh-deja package installs from the plugin side with dsh plugin --profile web add dsh-deja — the package reads what deja install wrote and contributes only what is missing. The DeepSeek Harness row asks for zstd, and session resume is marked unavailable there; handoff works by paste.
Does my session history leave the machine?
Per the README's FAQ, nothing leaves the machine unless you ask. Known secret shapes — AWS keys, api_key/token assignments, bearer tokens, bare JWTs, PEM blocks — are stripped as the index is built, and the README itself warns that pattern matching is not secret detection.
How is it different from the memory plugins already on this map?
Per its README's own framing: memory tools that record forward start empty, while deja starts full — it indexes the history the agents already wrote, including from before installation, and reads existing session files rather than maintaining a store of its own.
What does a recall cost at runtime?
No model calls: per its README, a recall is a lexical lookup against the local index with a ~0.4 ms median in-process, and the index updates incrementally, re-reading only session files that changed.
Official references