# IHMT Memory [Health: Active]

**Category:** 🧠 Knowledge & Memory  
**Repository:** https://github.com/gonzaroman/IHMT-MEMORY  
**GitHub Stars:** 35  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/ihmt-memory

## Description
Long-term memory for AI coding agents, stored as a tree of plain local files.

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `uvx` (confidence: high):

```json
"mcpServers": {
  "ihmt-memory": {
    "command": "uvx",
    "args": ["ihmt-memory"],
    "env": {
      "IHMT_HOME": ""
    }
  }
}
```

**Requires environment variables:** `IHMT_HOME` — the values above are empty placeholders; fill in real credentials before running (see the repository for what each one is for).

## Documentation

## What the IHMT Memory MCP server does

The IHMT Memory MCP server provides persistent context for AI coding agents. It stores memories on the local filesystem as readable text files arranged in a hierarchical tree, allowing information to continue across sessions, projects, models, and supported agents. Typical memory includes architectural decisions and their reasons, development-environment details, user preferences, project information, people, and corrections.

The memory is shared when multiple agents point to the same folder. Claude Code, Codex, and opencode are tested in the project documentation. Other clients that support local MCP servers may also work, but are not identified as tested. Each new installation starts with its own empty memory unless it is configured to use an existing folder.

## How it works

IHMT organizes raw text leaves beneath recursive JSON summaries, with a root JSON file at the top. A search starts at the root and follows relevant branches until it reaches matching leaves. This makes the number of files read depend mainly on tree depth rather than on the total number of stored entries. The project describes the retrieval cost as approximately beam times the logarithm of the number of memories, rather than a scan across all entries.

The IHMT Memory MCP server also preserves changing information. When a fact is replaced, the previous version remains as history and is marked `OUTDATED`; searches prefer the current version. Replacement links can be supplied when an agent corrects a prior memory. If a query is ambiguous, such as a name shared by multiple people, the system asks for clarification rather than selecting a likely match.

## Setup and configuration

Installation requires Python 3.10 or newer, git, and an AI agent that can run commands or edit files. Internet access is needed during installation to obtain the code and package, but the README says it is not needed afterward. macOS and Ubuntu are tested; Windows instructions exist but are not tested.

The repository can be cloned into `~/IHMT-MEMORY`, installed in a Python virtual environment, and started with `mcp_server.py`. The package can also be run through `uvx ihmt-memory`, or installed with `pipx`. By default, memory is kept in `~/.ihmt`; set `IHMT_HOME` to use another directory or an existing memory collection.

Each client needs a local MCP configuration and the usage-instruction template appropriate to that client. The documented configurations cover Claude Code, Codex, and opencode. After setup, start a new agent session so its memory tools are loaded. Chat-only browser assistants cannot perform the installation because they cannot run the required commands.

## Tools and capabilities

The available memory workflow supports searching existing information and saving information for later sessions. Agents can retrieve only the context relevant to a question, save durable decisions or corrections, associate a replacement with an older memory, and use current-versus-outdated status when information changes. The shared folder can be copied, backed up, or placed under version control because its contents are plain text and JSON.

## Limitations and notes

IHMT is local: it does not require a cloud account, database, or paid service, and its stored files are not sent elsewhere by the server. Retrieved memories still become model context when an agent uses them. The documentation reports tested support for Claude Code, Codex, and opencode; compatibility with other standard local-MCP clients is expected but not tested. Python 3.10 or newer is required, and Windows support remains untested.

_Full upstream README: https://allmcps.com/mcp/ihmt-memory/readme_