IHMT Gives AI Coding Agents Persistent Local Memory Through Plain Files

Summary

IHMT is an open-source, local memory system for AI coding agents. It is implemented as a standard stdio MCP server and is officially supported for Claude Code, with Codex and opencode tested; other MCP-compatible clients may work but are not yet tested. The system stores memories as UTF-8 text files and JSON metadata on disk, with no cloud service, database, account, or paid dependency beyond the agent itself. A recursive tree of leaves, branch summaries, and a root index lets searches descend through the hierarchy instead of scanning every stored item. The project says typical retrieval uses about 200–900 tokens whether the store contains 50 or 50,000 entries, though its token estimates are approximate. IHMT preserves older facts as history, marks them OUTDATED when a newer correction is saved, and asks for a clue when a query is ambiguous rather than guessing. It can share one memory folder across Claude Code, Codex, and opencode, or use separate project stores. The repository includes a Python 3.10+ CLI, a standard-library graphical interface, ingestion for code, narrative, temporal, and process documents, and optional Anthropic-based summarization with a heuristic offline fallback. Its MCP server exposes long-term memory, project indexing, and session scratch tools. ProjectIndex can synchronize code by checksum and return symbols rather than whole files, but the README reports mixed results: on a 55-file Spring Boot project it reduced selected lookup output compared with reading whole files, while a forced A/B test made sessions 42–71% more expensive and added about 460 tokens per conversation when the server was enabled. The project reports 187 tests, is tested on macOS and Ubuntu, and includes untested Windows instructions. It is released under the MIT license.