Pensieve 0.5.9 builds a local index of repositories, documents and previous agent conversations, then returns relevant passages with file and line references through the Model Context Protocol, or MCP. Coding agents can search for the meaning of code before guessing which strings to grep.
For Anthropic’s Claude Code, Codex and other coding clients, the proposed benefit is less exploratory reading. An agent can ask where a behavior lives and retrieve a narrower starting point, saving repeated searches for identifiers and file openings.
The Pensieve product page reports fewer tool calls, lower token use and faster lookups in the author’s benchmark. Those results deserve attention, though they measure a specific code-finding workload, not completed engineering tasks. They also come from Pensieve 0.5.6, while the advertised release is 0.5.9.
Meaning Search Meets Exact Matches
According to the project README, Pensieve embeds text using Qwen3-Embedding-0.6B and stores it in SQLite, with a trigram full-text index for exact matches and turbovec indexes for vector search.
Exact search works well when an agent knows an identifier or error message. Semantic search offers another route when the request describes behavior without naming its implementation, such as finding code responsible for retrying a failed payment.
Pensieve supports both approaches. Plain language searches by meaning; quoted phrases request exact matches. Queries can combine them, exclude terms with a minus sign, and filter by content type, extension or location.
The indexed sources extend beyond committed code:
- Local documents and folders, including Obsidian vaults and optionally Apple Notes.
- Git repositories, including uncommitted edits and worktrees.
- Claude Code, Codex and Qwen Code sessions.
The documentation says Pensieve watches these sources and reindexes changes, with worktrees tracked by branch. Agents can therefore search changes that do not yet exist in the main branch.
Past sessions add earlier investigations and decisions alongside their implementation. They also require judgment: a previous conversation may contain a rejected approach or an outdated explanation. Retrieval should lead the agent back to current code, without making historical discussion authoritative.
Scopes Keep Retrieval Attached to the Job
MCP supplies the connection between the index and the coding client. Pensieve supports stdio through and a local HTTP endpoint.
