Dream auto-extracts your session history from Claude Code, Codex, and Cursor into a unified local SQLite store. Agents recall your past context over MCP and lifecycle hooks — you never have to say "remember that" again. Single static Go binary. No API key. No cloud.
Three primitives. Plug into any coding agent.
Watches Claude Code, Codex, and Cursor session files as they're written. Extracts conversations, tool calls, and code edits — no copy-paste, no manual annotation, no MCP server you have to manage. Just runs.
All your agent history in one queryable file. Quantized ONNX all-MiniLM-L6-v2 embeddings live next to the rows; nothing leaves your machine. ~57 MiB of local assets, no network at query time.
Exposes recall as an MCP tool. Plugs into Claude Code lifecycle hooks so your past context surfaces automatically at session start. Drop-in for any agent that speaks MCP.
Measured on LongMemEval cleaned, 50 dev / 450 held-out, reproducible from the repo.
| Metric | Local only (no API key) |
+ LLM rerank (top-10) |
|---|---|---|
| R@5 (held-out 450) | 98.9% | 98.9% |
| R@10 (held-out 450) | 99.1% | 99.1% |
| R@1 (held-out 450) | 87.1% | 93.3% |
The 98.9% R@5 is the no-LLM result — what you get with make setup and no extra config.
To our knowledge it sits at or above the highest published held-out R@5 for any local-first memory system on this
dataset. Adding a Haiku rerank pass over the top-10 fused candidates lifts R@1 by +6.2pp without
changing R@5 — useful when the consuming agent reads only the #1 hit.
Install in one command. Runs offline. Free and open source.