What is deja-vu?
deja-vu (the command is deja) builds a local search index over the transcripts that Claude Code, Codex, Cursor, Gemini CLI, opencode and about 40 other coding agents already write to disk. Agents get that history back through a single MCP tool and through hooks that fire at session start, before a file edit and when a command fails; you get a terminal search screen, a local HTML export and commands such as deja blame, deja fix and deja wip.
It calls no language model and needs no account. Credentials are redacted as the index is built, and the only network traffic is updates, optional sync over SSH, a daily release check and an embeddings endpoint if you choose to add one for semantic search.
It is written in Go under the MIT licence, installs with a shell script, Homebrew, Scoop, winget, go install or npx, and ships plugins for several agents' own marketplaces. The repository was created in July 2026; v0.24.0 shipped on 5 November 2026.
Is deja-vu worth it?
deja-vu's central idea is the right one: your agents have been writing a full record of every session to disk all along, so a memory layer should index that record rather than ask the agent to decide what is worth remembering. The payoff is that recall works on day one, across agents. A fix worked out in Codex last month can surface in a Claude Code session today without anyone having written it down.
The numbers in the README are unusually well documented for a four-month-old project. It reports 97.2% recall at 5 on LongMemEval-S, 13.6 seconds to index 19,195 sessions, and a lookup median of 0.7 to 0.8 ms on a real store of 2,419 sessions. On a task the machine had already solved, runs with deja used 53,558 tokens against 126,222 with nothing wired, about 58% fewer. The single MCP tool costs 477 tokens of definitions a turn, against 8,283 for the largest of seven memory servers it measured. These are the maintainer's own benchmarks, but the drivers and corpora ship in the repository so anyone can rerun them.
Because it is lexical search with no model in the loop, it finds exact strings well (error messages, file paths, flags) and rephrased questions less well. Optional embeddings through Ollama, LM Studio or an OpenAI-compatible endpoint close that gap, at roughly 4 MB per thousand messages for a 1,024-dimension model; with a remote endpoint the redacted text and every query leave the machine.
The privacy story is better than most memory tools, with caveats worth reading. Redaction strips keys, tokens, JWTs and passwords from the index, but the source transcripts still hold them; the docs cite one machine with 84 secrets in 42 sessions, and deja secrets --scrub rewrites the originals only where it can. The index also keeps sessions an agent has since deleted, which is a feature for recovery and a liability on a shared or audited machine.
The pace is the other thing to weigh. Support for 43 agents, with formats that change when the agents update, means frequent releases and 20 open issues and pull requests. Against record-forward tools such as Mem0 or claude-mem, which start empty and hold only what an agent chose to save, deja trades curation for coverage, and its promote and remember commands add a curated layer on top when you want one.
Where it earns its place
A backend developer who uses Claude Code at work and Codex for side tasks
Stop the agent re-debugging a JWT rotation bug already fixed months ago
Run deja install --auto once; session-start hooks inject up to 2 KB of matching history from the current project, and the agent can call the deja tool in fix mode with the failing output. Cost is zero beyond a few hundred tokens a session.
A tech lead reviewing a change an agent made to a payments module
Find out why a line was written the way it was
deja blame path/to/file.go:120 names the commit that last changed the line and the agent session that wrote it, with what was decided; --git-note records the answer in refs/notes/deja for later reviewers.
A freelance developer whose Claude Code transcripts expire after 30 days
Keep and restore old sessions before the cleanup deletes them
Sessions indexed before cleanup stay searchable after the files are gone, and deja resume <id> --write-back puts one back for claude --resume. The index for a 19k-session store is about 185 MB.
Before you commit
- Setup
- curl the install script (macOS and Linux), brew install deja-vu, or Scoop or winget on Windows, then deja install --auto, which wires MCP and hooks into every detected agent and builds the first index in about 10 seconds. npx @vshulcz/deja-vu tries it without installing.
- Integrations
- One stdio MCP tool, deja, with modes recall, context, blame, fix, how, orient, remember and handoff; hooks for session start, pre-edit and failed commands; plugins for Codex, opencode, Claude Code, Gemini CLI, Zed, OpenClaw, Hermes and others; JSON output for scripts.
- Data and privacy
- Index in ~/.cache/deja and notes in ~/.local/share/deja on your machine. Network use is limited to updates, a daily release check (DEJA_OFFLINE=1 stops it), deja sync ssh and an embeddings endpoint you configure. MIT licence.
- Cost at scale
- Free. The index runs to about 4 to 10% of transcript size (232 MB for a 5.7 GB store); a cold build of that store took 71 seconds. Optional embeddings cost whatever your endpoint charges, or nothing with local Ollama.
- Limits
- Lexical search misses rephrased questions without embeddings; TRAE IDE chats are encrypted and cannot be read; some agents lack auto-recall or resume; 20 open issues and pull requests.
What’s good
- Searchable history from day one across about 43 agents
- No LLM, no account; lookups under 1 ms on a 2,419-session store
- One MCP tool at 477 tokens of definitions per turn
- Credentials redacted at index time; benchmarks reproducible from the repo
What’s not
- Lexical by default; paraphrased questions need an embeddings endpoint
- Source transcripts still hold secrets the index has redacted
- Keeps deleted sessions, which some teams will not want
- Fast-moving project created in July 2026; formats track 43 agents
Best for
Developers who switch between several coding agents and want each one to know what the others already fixed, without a hosted memory service.
deja-vu alternatives
deja-vu: common questions
- Is deja-vu worth it?
- Memory for coding agents that starts with everything you have already done, because it reads the transcripts instead of waiting for the agent to save notes. We rate it 4.2 out of 5.
- How much does deja-vu cost?
- Free under the MIT licence; no hosted or paid edition.
- What are the downsides of deja-vu?
- Lexical by default; paraphrased questions need an embeddings endpoint; source transcripts still hold secrets the index has redacted; keeps deleted sessions, which some teams will not want; fast-moving project created in July 2026; formats track 43 agents.
- Who is deja-vu best for?
- Developers who switch between several coding agents and want each one to know what the others already fixed, without a hosted memory service.
- How much does deja-vu cost for a team or in production?
- Free. The index runs to about 4 to 10% of transcript size (232 MB for a 5.7 GB store); a cold build of that store took 71 seconds. Optional embeddings cost whatever your endpoint charges, or nothing with local Ollama.
- Is deja-vu safe for company data?
- Index in ~/.cache/deja and notes in ~/.local/share/deja on your machine. Network use is limited to updates, a daily release check (DEJA_OFFLINE=1 stops it), deja sync ssh and an embeddings endpoint you configure. MIT licence.
- What are the best deja-vu alternatives?
- The closest alternatives we have reviewed are claude-mem and Mem0.
Answers checked Nov 5, 2026.
Make deja-vu? Show this rating on your site
