Pensieve is pitching local semantic indexing as a way to give software-development agents relevant context without repeatedly scanning entire repositories. The free, open-source Mac application searches documents, folders, Git repositories and past coding-agent sessions by meaning, exact text or a mixture of both.
For people, Pensieve provides a keyboard-activated search panel and filters for file type, repository and content category. For agents, it exposes the same index through the Model Context Protocol, returning selected passages with file and line references. The developer says it can connect to MCP clients including Claude Code, Codex, Cursor and other development tools through either standard input/output or HTTP.
The core promise is local operation. Indexing and search run on the user's Mac with an on-device embedding model and a local index. Users choose which folders, repositories, applications and sessions are included, and disabling a source removes its indexed material. The project says it excludes password-manager exports, keys, environment files and credentials. Optional AI-generated summaries are separate and can remain disabled.
Pensieve also tracks working copies rather than only committed code. It watches repository changes, includes uncommitted edits and follows worktrees by branch. Beyond file retrieval, it uses Git blame and recent history to map which contributors know particular parts of a codebase and to flag areas where knowledge may be concentrated among inactive contributors.
The project's benchmark claims measurable gains, though the results should be read as developer-run tests rather than an independent evaluation. Across 144 runs on held-out code-finding tasks in four real repositories, Pensieve reportedly found the correct file in 131 runs, compared with 123 using grep alone. The agent averaged 2.0 tool calls per lookup instead of 3.4, while the developer reports 16% fewer tokens, an 18% lower cost and a 21% speed improvement. Repository sizes ranged from 504 to 34,847 files.
The benchmark used Claude Sonnet through Claude Code and repeated 72 tasks twice. The project acknowledges that grep performed best for exact error messages, while semantic search did better on other forms of description. Its benchmark scripts are open source, allowing teams to rerun the comparison on their own code.
Pensieve addresses a practical bottleneck in agent-assisted development: finding a small amount of useful context without paying to read large volumes of irrelevant text. Its local design may appeal to teams wary of sending repositories to a search service. Whether the reported savings generalize will depend on codebase structure, query types and agent behavior, but the reproducible benchmark gives prospective users a concrete starting point for evaluation.
