Adaptive Memory Graph

An MCP server plugin that gives Claude persistent, intelligent memory across sessions. It stores knowledge as weighted, interconnected nodes in a graph that evolves through conversation — nodes that get used gain weight, unused ones decay and eventually archive.

Works with Claude Code and Claude Desktop.

Features

• Weighted memory nodes — Important memories stay prominent; stale ones fade
• Cross-domain connections — Link related knowledge across topics
• Time-based decay — Graph self-prunes so only relevant memories persist
• Encrypted storage — AES-256-GCM encryption with macOS Keychain key storage
• Session logging — Tracks which memories were accessed and how they were received
• Domain organization — Nodes organized by domain (e.g. health_and_safety, personal, ideas_and_projects)
• Chat history ingestion — Review and extract knowledge from past Claude Code sessions

Installation

pip install adaptive-memory-graph

Or with uv:

uv pip install adaptive-memory-graph

Setup

Claude Code

claude mcp add adaptive-memory-graph -s user -- amg-server

Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "adaptive-memory-graph": {
      "command": "amg-server"
    }
  }
}

Config file location:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

Tools

Tool	Description
amg_load_index	Load lightweight graph index at session start
amg_expand_branch	Fetch full node content when contextually relevant
amg_get_connected_nodes	Find related nodes across domains
amg_log_session	Log session summary at conversation end
amg_update_graph	Process pending logs and apply weight decay
amg_export_report	Generate human-readable graph summary
amg_manual_adjust	Boost, decay, archive, or delete nodes
amg_add_node	Add new nodes to the graph
amg_search_nodes	Search nodes by title, summary, tags, or content
amg_list_chat_sessions	List available Claude Code chat sessions for review
amg_read_chat_session	Read a chat session's conversation content

How It Works

1. Session start — Claude calls amg_load_index to get a lightweight summary of your memory graph
2. During conversation — If a topic is relevant, Claude expands specific nodes for deeper context
3. Session end — Claude silently logs which nodes were accessed and suggests new ones
4. Between sessions — Weight decay runs, archiving memories that haven't been useful

Nodes are stored as encrypted JSON on disk (~/.amg/graph.json.enc). The encryption key is stored in your macOS Keychain.

Requirements

• Python 3.10+
• macOS (for Keychain-based encryption key storage)

License

MIT