deja-vu: one local memory index for forty-two coding agents, no LLM involved
deja-vu indexes the session history your coding agents already write to disk and serves the relevant piece to whichever agent is open. MIT license, one Go binary, no model calls. The benchmark claims are the maintainer's own; the privacy design reads well, with caveats.
Facts checked October 16, 2026
deja-vu is an MIT-licensed Go tool that indexes the session history Claude Code, Codex, Cursor and 39 more agents already keep on disk, then feeds recall into whichever agent you open, through MCP, hooks or a terminal UI. No model, no embeddings, no server. I read the README and LICENSE; I did not install or run it. Repository
What it does
Every major coding agent already writes its sessions to disk. deja-vu indexes all of that history, months back, into one local index, and all forty-two supported agents read the same one. A fix found in a Codex session can surface in Claude Code the next day. Recall can arrive unasked: at session start and after a command fails, through agent hooks. Running deja with no arguments opens a terminal search over everything, with filters by agent, date and project.
It also records what each turn touched: files opened, commands with exit status, the exact spans edits replaced. A promote command lets you mark a decision as rejected, and later hits say it was tried and dropped. Sessions their agent deleted stay in a Deleted tab unless you purge them.
Setup, from the README
On macOS and Linux the documented install is two commands: fetch install.sh from the repository and pipe it to sh, which drops the binary in ~/.local/bin, then run deja install --auto, which wires MCP recall into every agent it finds and builds the first index. Alternatives: brew install deja-vu, go install, npx @vshulcz/deja-vu for a no-install try, Scoop or winget on Windows, and plugin bundles for Claude Code, Cursor, Qwen and OpenClaw. The binary alone is a complete search install; the install subcommand is only the wiring. deja doctor reports what is wired and what is missing.
As always with curl-to-shell installs, read the script first. It fetches a binary and runs it with your user rights.
The claims, labeled
The maintainer publishes his own measurements in the repo: 97.2% recall at 5 on LongMemEval-S against 96.6% for MemPalace and 95.2% for agentmemory, half the tokens of agentmemory on one repeated task, and 17.6 seconds to index 19,195 sessions against 72 seconds for the next of seven tools. The corpus, drivers and scoring are in the repository, so the rows are re-runnable, but they are still the maintainer's numbers on his machine, not an independent evaluation. Popularity is not quality either way: 1,196 stars and 115 forks as I checked today.
Maintenance and community
The repository was created on July 14, 2026. The main contributor, vshulcz, accounts for most of the 2,100-plus commits; a handful of others contribute smaller pieces, plus dependabot. The commit log shows activity on October 14, 15 and 16, the day I checked. Sixteen issues were open. Commit log
Privacy and security read
The design is local: no model, no embeddings, no server, and the README says keys and tokens are stripped while the index is built. That is the right shape for a tool that reads everything your agents did. The caveats are real though. The index is a second, searchable copy of your whole coding history, so it deserves the same protection as the sessions themselves. The Deleted tab keeps sessions you may have wanted gone. Auto-recall injects old session text into new agent contexts, so a poisoned or sensitive old session can reach a future agent without you asking. And the default install is a piped shell script, a trust decision worth making consciously.
Why it matters
Agent memory products are usually per-agent and usually cloud-backed. deja-vu makes the opposite bet in both directions: reuse the history already on disk, and never call a model to do it. If the recall claims hold up under independent testing, that is a cheap and private way to stop re-debugging what you already fixed. If you want to try it without installing, npx @vshulcz/deja-vu followed by a query runs it once.
License
MIT, per the LICENSE file in the repository. Copyright Vladislav Shulcz, 2026. LICENSE file
Try it, then a review checklist before wiring it everywhere
npx @vshulcz/deja-vu "connection pool exhausted" # one run, nothing installed
# my review checklist before wiring it into every agent:
[ ] read install.sh before piping anything to sh
[ ] run deja doctor to see every MCP target it would touch
[ ] decide about the Deleted tab: it keeps sessions the agents deleted
[ ] remember auto-recall feeds old sessions into new agent context