deja-vu

Your agent is about to re-debug something you fixed eight months ago, in a different agent than the one that fixed it. deja-vu, built by Vladislav Shulcz and released in mid-2026, is aimed squarely at that failure mode: 34 coding-agent harnesses (Claude Code, Codex, Cursor, aider, Zed and more) already write every conversation to local session files, and almost nothing reads across them. deja-vu indexes those files where they already live and hands the right one back to whichever agent asks, with no capture step, no embedding model and no API key required for the core path; the transcripts themselves are the memory.

A single Go binary parses each agent's own session format (JSONL for Claude Code and Codex, SQLite for Cursor) into a local inverted index, stripping credentials as it goes: AWS keys, bearer tokens, JWTs, PEM blocks and even a password stated in prose. Recall then runs as a lexical lookup, reported at 0.7-0.8 ms median on a real 1.9 GB, 2,419-session corpus, and it fires automatically at session start, before a tool call, and after a failure, not only when explicitly asked. An MCP server exposes the same index to any agent that supports it, and a deja stats command turns a year of sessions into a shareable terminal card or SVG.

It fits developers who already bounce between several coding agents on one machine and lose context every time they switch; the README's own pitch is 'solve it in Codex, Claude remembers.' Because indexing works retroactively over history written before deja-vu was even installed, there is no cold-start period, and because the core path needs no embeddings or LLM call, it works entirely offline with predictable, near-zero latency. The project also documents an explicit security model: what's redacted, what isn't (raw command output in the original transcripts is untouched), and how to exclude a project or forget a session.

The headline figures (88.1% hit@1 on LongMemEval-S, 70.5% on LoCoMo, the token-savings comparisons against full-history and naive grep) are the project's own benchmark harness, not an independent evaluation, though both harnesses ship in the repository and run on public datasets in minutes, which makes them unusually easy to check yourself. Windows gets real binaries and CI coverage but is explicitly called out as less battle-tested than macOS and Linux. And while the core path is local-first, optional semantic recall will send redacted, truncated indexed text to whatever embedding endpoint you configure, local by default but a real setting to double-check if privacy is the point.

deja-vu is a strong pick for anyone working across multiple coding agents who wants cross-agent memory without standing up a memory platform or paying for embeddings: it is free, MIT-licensed, actively maintained (33 contributors, near-weekly releases), and its performance claims come with the means to verify them yourself. The honest caveats are Windows maturity and the fact that every number quoted is self-reported, however reproducible. For a single-agent workflow, or if you specifically need semantic rather than lexical recall as the primary path, it is worth checking the comparison guide before assuming it is a drop-in fit.