MemoryChat MCP Server

Obsidian-vault-backed long-term memory for AI coding agents.

An MCP server that gives any MCP-compatible AI agent (Claude Code, Claude Desktop, Cursor, opencode, goose) persistent, human-editable memory stored as plain markdown files in an Obsidian vault.

Why

• Human-editable memory. When the AI remembers something wrong, open the markdown file in Obsidian and fix it. No vector database mystique, no opaque embeddings — just files.
• Obsidian graph for free. Memories cross-reference each other with wiki-links. Obsidian's graph view shows your AI's knowledge network.
• BYO-LLM. The server stores and retrieves notes. It doesn't call any LLM API. Use it with whatever model you already pay for.
• 5 memory types. User facts, feedback/corrections, project context, reference links, and session handoffs — mirroring the mental model of Claude Code's built-in memory.

Install

Requires Python 3.11+ and uv.

git clone https://github.com/dhawal/memorychat-mcp.git
cd memorychat-mcp
uv sync

Configure

Create ~/.config/memorychat/config.toml:

[vault]
path = "/path/to/your/obsidian/vault"
memory_subdir = "_memory"

[search]
backend = "keyword"

[dedup]
similarity_threshold = 0.85

[server]
log_level = "INFO"
log_file = "~/.local/state/memorychat/server.log"

Create the memory directory structure in your vault:

cd /path/to/your/obsidian/vault
mkdir -p _memory/user _memory/feedback _memory/project _memory/reference _memory/episodic

Use with Claude Code

Add to ~/.claude/mcp.json:

{
  "mcpServers": {
    "memorychat": {
      "command": "uv",
      "args": [
        "--directory",
        "/path/to/memorychat-mcp",
        "run",
        "python",
        "-m",
        "memorychat_mcp"
      ],
      "env": {
        "MEMORYCHAT_CONFIG": "~/.config/memorychat/config.toml"
      }
    }
  }
}

Restart Claude Code. The server provides 10 tools:

Tools

Tool	Purpose
memory_search	Find memories by keyword query with weighted scoring
memory_recall	Fetch a single memory by UUID (full body)
memory_write	Create a new memory note (with dedup check)
memory_update	Update an existing memory (partial updates)
memory_delete	Move a memory to trash (recoverable)
memory_list	List memories by type/project
session_start	Load core memory context for a new session
session_handoff	Write a structured session handoff note
memory_link	Create bidirectional links between memories
memory_index_rebuild	Regenerate the INDEX.md from current notes

Memory Types

Type	Purpose	Lifetime
user	Profile facts (role, expertise, preferences)	Long — survives projects
feedback	Corrections ("don't do X", "always do Y")	Long — survives projects
project	Project-specific context (goals, blockers)	Medium — lives with project
reference	Pointers to resources (URLs, paths)	Long, until resource moves
episodic	Session handoff notes	Long — append-only history

Example

You: Remember that I prefer using type hints in all Python code.

Agent: [calls memory_write(type="feedback", name="Always use type hints", ...)]

--- next session ---

Agent: [calls session_start(project="memorychat")]
Agent: I see from your preferences that you want type hints in all Python code.

The memory file at _memory/feedback/always-use-type-hints.md:

---
id: 7e3b1a4c-...
name: Always use type hints
description: User insists on type hints in all Python functions
type: feedback
created: 2026-04-13T18:42:11Z
updated: 2026-04-13T18:42:11Z
project: null
tags: [python, code-style]
links: []
sessions: []
---

The user has stated that every Python function must include type hints.
They consider untyped code to be incomplete.

Open it in Obsidian, edit it, and the next session picks up your changes.

Writing a high-confidence memory

When the user explicitly says "remember this," set confidence to "high":

{
  "tool": "memory_write",
  "arguments": {
    "name": "Prefers dark mode",
    "description": "User prefers dark themes across all apps",
    "type": "user",
    "body": "User prefers dark themes",
    "confidence": "high"
  }
}

High-confidence memories rank above medium and low-confidence ones in search results, regardless of recency. If confidence is omitted, it defaults to "medium".

Development

# Run tests
uv run pytest tests/ -v

# Lint
uv run ruff check src/ tests/

# Type check
uv run mypy --strict src/

# Run the server directly (blocks on stdin)
uv run python -m memorychat_mcp

License

MIT