Your agents already fixed this. deja reminds them.

What it does

deja-vu is a single zero-dependency binary that reads the conversation histories AI coding agents already write to disk—Claude Code, Codex, Cursor, aider, and others—and builds a searchable index from them. It works retroactively, pulling in months of past sessions before you even installed it, then serves relevant memories back to any agent via an MCP server or automatic session-start hooks. The goal is simple: when you ask an agent a question it already answered eight months ago, it answers from memory instead of starting over.

The interesting bit

The retrieval runs without touching an LLM, embeddings, or any API key, yet the in-repo benchmark harness reports 84.9% top-1 accuracy on LongMemEval-S. It searches roughly 3.3 GB in about a millisecond. Instead of trusting a cloud memory service, you sync append-only history between machines over SSH, and the index redacts API keys and JWTs at ingest time.

Key highlights

• Retroactive indexing of existing agent logs across seventeen supported harnesses, including Claude Code, Codex, Cursor, and Gemini CLI.
• ~1 ms median search over gigabytes of history with no LLM calls or embeddings required for core retrieval.
• Promote and reject sessions: curated notes outrank raw transcripts, and decisions you reversed surface as “tried and rejected” with the reason and date.
• deja sync moves memory between machines over SSH; deja handoff packages live context to switch agents mid-task.
• Local-only by default: a zero-dependency Go binary with an optional HTML view and built-in redaction of secrets.

Caveats

• First-class support varies by agent: Cursor is skipped for guidance files, Copilot drops hook context, aider lacks MCP and hooks entirely, and Grok only reads user-level guidance when no project-level file exists.
• On Windows, MCP server registration requires a shell wrapper rather than direct stdio.

Verdict If you bounce between Claude Code, Codex, and a dozen other agents and are tired of re-explaining the same architecture decisions, deja-vu is worth the disk space. If you only use one agent in a single project with no history, it has nothing to index and little to offer.