Callimachus

Every AI coding session solves something. Almost none of that solution survives into the next session, because there’s no index — just a flat list of threads scattered across five tools and three weeks. Callimachus runs quietly in the background and indexes those conversations as they’re written, building a local SQLite + on-device embedding index that you can query by keyword, by meaning, by file path, by project, or by recency. The result is a searchable catalogue of your own debugging history, not a cloud sync or another SaaS subscription.

The differentiating layer is hybrid search: SQLite FTS5 keyword matching fused with on-device semantic similarity. The docs describe this as letting a vague memory — ‘that vector index migration’ — surface the correct thread even when the exact words are gone. Subagent transcripts are indexed separately and stay collapsed until you ask for them. A ‘cited answer’ feature runs retrieval-augmented generation over your own session history, returning a synthesized response with [thread N] citations back to the source threads — though this requires an LLM engine, either local Ollama or a cloud API key.

Callimachus exposes the same index through four surfaces: a native desktop app for macOS, Windows, and Linux; a VS Code and Cursor sidebar extension; a `cal` CLI for terminal-based search and piping context into other tools; and an MCP server that lets any MCP-compatible agent query your history on demand. The MCP path is the sharpest edge for teams already running agent workflows — an agent can pull your past decisions rather than starting cold. The hard constraint is the supported tool list: eleven tools are indexed at v0.6.1. If your team’s primary tool is not on that list, no indexing occurs without a manual import path the vendor page does not describe.