Recoleta
GitHub
Source note

Show HN: Callimachus – Local search across your AI coding-agent history

Authors
arishaller
Source
Hacker News
Published
Collection
Software Intelligence
Code IntelligenceCoding AgentsLocal SearchDeveloper ToolsAgent Memory
Open GitHubMarkdown source

Summary

Callimachus is a local app that indexes AI coding-agent conversations and makes them searchable from desktop, CLI, VS Code/Cursor, and MCP clients. It targets developers who use several coding agents and need prior decisions, TODOs, file mentions, and transcripts without sending the index to a cloud service.

Problem

  • AI coding-agent work is split across many tools, so past decisions, fixes, and TODOs are hard to find during later coding sessions.
  • Lost thread history can cause repeated work, inconsistent project decisions, and weak context when switching between Claude Code, Codex, Cursor, Gemini CLI, and other tools.
  • The tool matters because agent memory helps only when developers and agents can retrieve it inside the editor, terminal, or agent session where work happens.

Approach

  • It imports conversations from 11 coding-agent sources into one local SQLite database.
  • Search combines SQLite FTS5/BM25 keyword ranking with on-device vector search through sqlite-vec, then merges rankings with Reciprocal Rank Fusion.
  • A file-mention index maps paths to threads, so a query such as file:embed/mod.rs can find sessions that touched that file.
  • Optional LLM passes extract decisions, gotchas, TODOs, summaries, conflicts, and cited answers over prior threads.
  • The same index is exposed through a desktop app, cal CLI, VS Code/Cursor extension, provider-agnostic chat, and an MCP server that lets agents read and write project memory.

Results

  • Supports 11 sources: Claude Code, Codex, Cursor, Gemini CLI, Qwen Code, Goose, OpenCode, Continue, Cline, Roo Code, and Kilo Code.
  • Ships 16 MCP tools, including thread search, current-project search, file-to-thread lookup, cited history Q&A, decision recall, gotcha recall, and memory writes.
  • The CLI exposes 21 commands, including search, recent, export, ask, files, memory, done, remember, agents, and hook.
  • The semantic index uses bge-small-en-v1.5 embeddings with 384 dimensions and sqlite-vec KNN running locally.
  • The author reports indexing a Claude corpus of about 90,000 messages in about 25 seconds, with later passes skipping unchanged files.
  • No benchmarked retrieval accuracy, baseline comparison, or user-study result is reported in the provided text.

Link