IHMT — Infinite Hierarchical Memory Tree

IHMT (Infinite Hierarchical Memory Tree) is an open-source, MIT-licensed long-term memory system for AI coding agents, published by gonzaroman. It runs as a stdio MCP server that lets agents such as Claude Code, Codex and opencode save decisions, preferences and corrections once and retrieve them in later sessions and across projects.

What It Is

IHMT stores memory locally as a recursive tree of plain files: raw text leaves at the bottom, JSON summaries above them, and a single root.json trunk. Searches walk the tree from root to branch to leaf rather than scanning a flat index, so the README says retrieval typically costs 200–900 tokens regardless of memory size. It needs no vector database, server or account, and the core uses only Python 3.10+ and the standard library.

How Memory Is Saved and Retrieved

The MCP server exposes nine tools in three families: long-term memory (search_memory, save_memory, mark_outdated), project indexes (project_map, find_code, read_file) and session scratch memory (note, recall, digest_output). Ingestion detects the type and domain of each document and splits it without cutting logical blocks, using code-, narrative-, date- and process-aware chunkers. Ranking uses BM25 and, when a query is ambiguous, the agent is asked for a clue instead of guessing.

Time Awareness

IHMT records dated facts such as location, workplace and stack. When a value changes, the older memory is kept as history, flagged OUTDATED, and searches return the current one first. Corrections can also be linked explicitly with the replaces parameter.

Setup Path

An agent can install it from a prompt that points to INSTALL.md, or users can install manually from a git clone or via uvx from the ihmt-memory PyPI package. Claude Code is officially supported; Codex and opencode are tested, and other MCP clients should work but are untested. A local graphical interface and a CLI are also included. The README states that matching is lexical rather than semantic and that writes assume a single writer.