vshulcz/deja-vu

deja-vu, or deja, is a command line tool written in Go that gives your AI coding assistants a memory. Tools like Claude Code, Codex, Cursor, and similar coding agents already write every conversation to local files on your computer, building up gigabytes of past debugging sessions and design decisions that you normally cannot search. deja reads those existing histories and turns them into a searchable memory layer, working retroactively on months of sessions from before you even installed it. At its core, deja lets you search across all your past agent sessions in milliseconds, even over gigabytes of history. It also exposes a recall tool through MCP, a protocol coding agents can call directly, so an agent can check whether a problem was already solved in a previous session instead of re debugging it from scratch, even if that earlier session happened in a different tool. There is a session start hook that can automatically surface relevant memory before you even ask, ranked by which files you are currently working on. Beyond search, deja can sync your memory between machines over SSH without any cloud service involved, hand off context from one agent to another mid task, distill a messy session into a clean curated note with a status like accepted or stale, and share a sanitized digest of a session with a colleague. A policy file controls trust scopes, deciding exactly what memory is allowed to activate in search, MCP, or auto recall, and whether it applies locally, to imported data, or per remote peer. Privacy is treated carefully. API keys, JWTs, and private keys are stripped out automatically at the time content is indexed. A forget command removes matching sessions and records a tombstone so they cannot be silently restored later, and exclusion patterns let you keep entire projects out of the index in the first place. deja ships as a single dependency free binary with no models to download and no background services to run, nothing leaves your machine unless you explicitly sync or share it. It can be installed through a shell script, Go, npm, or Homebrew, and it automatically wires itself into whichever supported coding agents it finds on your system. The project is released under the MIT license.