Skip to content

Summary

Indexes the coding-agent sessions already sitting on your disk — months of history from before you installed it — and serves them back over MCP. 3.5 GB searched in ~1.5 ms, no LLM, no embeddings.

Features

  • Indexes agent sessions already on disk — history from before you installed it
  • ~1.5 ms median search across 3.5 GB
  • 84.9% hit@1 on LongMemEval-S, 69.8% on LoCoMo — harnesses included in the repo
  • No LLM, no embeddings, no API key
  • Reads across seventeen agent harnesses
  • Session-start auto-recall where the harness supports it
  • One zero-dependency binary; moves between machines over SSH
  • Per-platform .mcpb bundles for desktop MCP clients

Installation

Set up this MCP server in your favorite AI agent — copy a ready-made configuration below.

Any MCP-compatible agent

Most agents (Claude, Cursor, Windsurf, VS Code, and more) read a standard mcpServers configuration.

  1. Open your agent's MCP configuration file.
  2. Merge the snippet below into it.
  3. Restart the agent — the "deja-vu" tools will be available.
{
  "mcpServers": {
    "deja-vu": {
      "command": "npx",
      "args": [
        "-y",
        "@vshulcz/deja-vu",
        "mcp"
      ]
    }
  }
}

Description

Every other memory tool starts empty and records forward, which means it is useless for the first few months. deja-vu starts full. It indexes the session transcripts your coding agents have already written to disk — months of history from before you installed it — and hands them back to any agent over MCP.

The numbers

Median search over 3.5 GB is about 1.5 ms. Retrieval accuracy is 84.9% hit@1 on LongMemEval-S and 69.8% on LoCoMo, with zero LLM calls, zero embeddings and zero API keys. Both benchmark harnesses ship in the repository and run on the public datasets in minutes, so the claims are checkable rather than asserted.

What it recalls

The pitch is a question asked twice: without deja-vu the agent has no record of it; with deja-vu it answers using a decision made eight months earlier — and nobody ran a search, the agent called deja itself. It reads across seventeen agent harnesses, so a decision recorded in Claude Code is available in Codex or OpenCode later.

Install and wiring

One zero-dependency Go binary, fully local:

curl -fsSL https://raw.githubusercontent.com/vshulcz/deja-vu/main/install.sh | sh

Also available through go install, npx @vshulcz/deja-vu, and Homebrew (brew install vshulcz/tap/deja-vu). Desktop clients that take MCP bundles can use the per-platform .mcpb from each release, which carries the binary so there is nothing else to install.

deja install --all wires MCP recall into every agent it finds on the machine, editing configs and keeping a .bak; deja install --auto adds session-start auto-recall where the harness supports it. Install also builds the index from your existing history. It moves with you between machines over SSH, and ships shell completions for bash, zsh and fish.

MIT licensed, with an OpenSSF Scorecard badge on the repository.

Related MCP Servers

MCP: Engram

by Gentleman Programming

New

Persistent memory for coding agents as a single Go binary — SQLite and FTS5 behind MCP, an HTTP API, a CLI and a TUI, with no Node, Python or Docker required.

MCP: agent-device

by Callstack

New

Give a coding agent a live feedback loop on real mobile, TV and desktop apps — inspect, tap, type and capture evidence on iOS, Android and HarmonyOS over MCP or a CLI.

New

The official W&B MCP server: query experiment runs, Weave LLM traces, artifacts and registries in natural language, and write findings back as a W&B report.

Featured

Replicate's official MCP server: search thousands of hosted models, read their schemas, and run predictions on image, video, audio and language models from inside an agent.

Browse all MCP servers →