vshulcz / deja-vu
A local Go binary that gives 34+ AI coding agents a shared memory by indexing the session history already on your disk — no LLM, no embeddings, no new capture step.
星标趋势
AI 分析
项目摘要
deja-vu is a local Go binary that builds a shared memory layer for 34+ AI coding agents (Claude Code, Codex, Cursor, Copilot CLI, etc.) by indexing the session history already on disk. It requires no LLM, no embeddings, and no new capture step — agents automatically recall relevant past fixes and decisions at session start, on every prompt, before edits, and after failures. It strips keys/tokens during indexing and serves results in milliseconds over gigabytes of history.
为什么值得关注
deja-vu solves a real pain point — AI coding agents repeatedly re-debug problems you solved months ago in a different tool — with an elegantly simple approach: index what's already on disk, no model needed. The cross-agent memory sharing (a fix found in Claude Code comes back in Cursor) is genuinely novel, and the benchmark numbers (88.1% hit@1 on LongMemEval-S, 58% token reduction) are compelling. The project is also remarkably active with 58 releases in 6 months and 33 contributors.
优势
- Zero-config approach: indexes existing session history with no new capture step, no LLM, no embeddings — just a local Go binary
- Cross-agent memory sharing: a fix found in one agent (e.g., Claude Code) is recalled in any other (e.g., Cursor, Codex), breaking down tool silos
- Strong benchmark results: 88.1% hit@1 on LongMemEval-S, 70.5% on LoCoMo, 58% token reduction, millisecond lookups over GBs of history
- Security-conscious: strips keys and tokens during indexing, with a documented security model
- Very active development: 58 releases in 6 months, 33 contributors, CI, Docker support, MCP server integration
局限性
- No automated tests listed in the repository, which is concerning for a tool that indexes sensitive session data
- Relies on parsing session history formats from 34+ agents — any agent that changes its storage format could break indexing silently
- The high release cadence (58 in 6 months) suggests the project is still in rapid iteration, which may mean breaking changes or instability for production users
使用场景
- Preventing AI coding agents from re-debugging previously solved problems across different tools (Claude Code, Cursor, Codex, etc.)
- Reducing token costs by giving agents memory of past decisions and fixes instead of re-deriving them
- Building a unified memory layer for teams using multiple AI coding agents who want shared context