AI news · October 13, 2026
Pensieve gives coding agents local search across files and past sessions
Pensieve reports 43% fewer tool calls and 18% lower lookup cost in a 144-run benchmark; builders can test local context search before adding a hosted retrieval service.
- reported fewer tool calls
- 43%
- reported lower lookup cost
- 18%
- agent runs in the benchmark
- 144
Pensieve is a free, open-source Mac tool that searches files, repositories and AI agent sessions locally, then returns the few passages that matter through an MCP server. The product connects to Claude Code, Codex, Cursor, Gemini CLI, Windsurf, Claude Desktop and VS Code. It also adds git-blame ownership and knowledge-gap views, so an agent can find both the right code and the person who knows it. The project says indexing and search run on the Mac, with no account or cloud service required for the core features.
The maker's benchmark compared agent lookups with normal grep, glob and read calls across four repositories. Pensieve reports 131 of 144 tasks solved versus 123 of 144, 2.0 tool calls per lookup versus 3.4, 18% lower cost and 21% lower time. These are project benchmarks, not independent results, and local indexing has setup and storage cost. Still, the experiment is easy to repeat: connect it to one repository, compare the top result and total reads, and keep it only if it improves real work.
What you can do with it
Open the project site, install Pensieve on a test Mac and index one repository plus one agent session. Run the same ten questions with and without the MCP server, then record correct-file rate, tool calls, latency and index cost. Keep source selection narrow until you understand what is stored locally.
Our take
Pensieve attacks a real cost in agent work: reading too much context before finding the useful file. Its local-first design and MCP connection make a small trial possible, while the ownership view adds a human handoff benefit. The benchmark is maker-run, so the decision should come from your own repository and question set.
Links Try itSourceHN discussion
Source: Pensieve / Hacker News — Made With Models writes the brief; the reporting is theirs.