AI KNOWLEDGE DESK

Models · entities · concepts · comparisons · practical tools

GETLLMS.ORG
ToolCoding agent memory

Deja Vu

Deja Vu is an open-source Go command-line tool that indexes local coding-agent session histories and lets users search or recall prior work across 20 documented harnesses, including Claude Code, Codex, OpenCode, Cursor, DeepSeek Harness, Oh My Pi, Zed, and Copilot CLI. It can expose recall through MCP, inject relevant context at session start where hooks exist, and optionally keep multiple machines synchronized.

Why it matters

Coding-agent sessions often repeat investigations because useful context is buried in old transcripts. Deja Vu makes those histories searchable without requiring a hosted memory service, but users still need to understand local-data boundaries, secret redaction, optional sharing or sync, and platform maturity.

Source-backed summary

The vshulcz/deja-vu repository and v0.18.0 release document 20 harnesses, timer-driven cross-machine sync, source-machine provenance, stale-sync reporting, new Zed, Oh My Pi, and DeepSeek Harness integrations, Qwen resume support, and more answer-centered recall. The release also adds reciprocal-rank fusion and performance and correctness fixes. These are publisher-reported product changes, not independent security, recall-quality, or benchmark proof.

Primary use cases
  • Search earlier Claude Code, Codex, and OpenCode sessions.
  • Recall prior debugging or implementation decisions at session start.
  • Expose local agent-history search through MCP.
  • Share a deliberately selected session after reviewing redaction and network boundaries.
How local recall works

Deja Vu scans supported local session histories, builds a lexical index, and lets users search prior conversations from the terminal. Version 0.18.0 expands the documented set to 20 harnesses, reads DeepSeek Harness sessions, adds Zed and Oh My Pi integrations, resumes Qwen sessions, and records Grok spawn trees as related work instead of unrelated transcripts.

What changed in recall and sync

Version 0.18.0 can keep machines aligned on a timer, records which machine a memory came from, and reports stale synchronization. Recall now quotes more surrounding answer context, raises the per-prompt recall budget to 1,536 bytes, fuses two ranking views with reciprocal-rank fusion, and avoids repeating an already shown block. These features improve provenance and selection, but teams should still test recall quality on their own histories.

Privacy and sharing are separate modes

The repository says indexing and search stay local and include secret redaction. Version 0.15.7 says the recall kill switch and policy now bind every recall path. Sharing, statistics, and sync are explicit features with different data boundaries, so teams should review each command before enabling it and should not equate local search with every optional networked workflow.

Adoption limits to check

Deja Vu documents history discovery across 16 harnesses, but MCP and session-start hook support differ by client. It ships Windows builds while describing macOS and Linux as the battle-tested platforms. Teams should verify session-path coverage, repository exclusion needs, redaction behavior, index storage, sync boundaries, and whether lexical recall is sufficient for their history volume.

Deja Vu FAQ

Common questions about Deja Vu.

Does Deja Vu upload coding-agent sessions?+

The repository says indexing and search run locally without network access. Optional sharing, statistics, or sync commands have separate network boundaries, so review the exact command and redaction behavior before using them with sensitive histories.

Which coding agents does Deja Vu support?+

The project documents 20 local history formats: Claude Code, Cline, Codex CLI, OpenCode, aider, Gemini CLI, Cursor, Antigravity, Grok Build, Hermes, Goose, Qwen Code, Kimi Code, pi, Oh My Pi, OpenClaw, Copilot CLI, Roo Code, DeepSeek Harness, and Zed. MCP, automatic recall, resume, and handoff still vary by harness.

Does Deja Vu work on Windows?+

The project publishes Windows builds, while describing macOS and Linux as the battle-tested platforms. Windows users should verify local session discovery, path handling, MCP startup, and redaction before relying on it for routine recall.