agentic-memory-mcp

Agentic Memory Management via MCP — Knowledge Graph for AI agents.

A Model Context Protocol (MCP) server that gives AI agents persistent, structured, and queryable memory via a Knowledge Graph. Designed to be used standalone, embedded in any agentic framework, or served as an MCP tool server.

Architecture

agentic_memory/
├── core/
│   ├── graph.py      # KnowledgeGraph — nodes, edges, JSON persistence
│   ├── score.py      # Score & ScoreStore — reputation / relevance scoring
│   └── prune.py      # AutoPruner — maintenance & auto-pruning
├── mcp/
│   └── server.py     # MCP server (stdio transport)
├── cli/
│   └── main.py       # Click CLI (memory-cli command)
├── api.py            # MemorySkills — Python API (standalone use)
└── hermes/           # Hermes Agent plugin integration

Persistence: JSON files (memory_graph.json, memory_scores.json) — no external database required.

Installation

# From PyPI
pip install agentic-memory-mcp

# Development
git clone https://github.com/alexandrerodenas/agentic-memory-mcp.git
cd agentic-memory-mcp
uv sync

Quick Start

CLI

# Add a memory entry
memory-cli --graph memory.json node add --id alice --label Person --content "Alice lives in Paris"

# Search
memory-cli retrieve --query "Paris"

# Link two entries
memory-cli --graph memory.json edge add --id e1 --source alice --target paris --label lives_in

# Get stats
memory-cli --graph memory.json stats

# Prune old entries
memory-cli --graph memory.json prune --max-nodes 500 --strategy oldest

MCP Server

# Start the MCP server (stdio)
memory-mcp

# Or with custom paths
MEMORY_GRAPH_PATH=/data/memory.json MEMORY_SCORES_PATH=/data/scores.json memory-mcp

Then connect any MCP-compatible AI agent (Claude Code, OpenCode, etc.) to use the tools:

• memory_node_add — Add a fact or entity
• memory_node_update — Update an entry
• memory_node_delete — Delete an entry
• memory_node_get — Retrieve by ID
• memory_retrieve — Token-optimized top-N retrieval by score (with optional text filter)
• memory_edge_add — Create a relationship
• memory_corroborate — Reinforce a memory (increases score)
• memory_prune — Manual pruning
• memory_stats — Graph statistics

Hermes Agent Plugin

# Install
hermes plugin install agentic-memory-mcp

# Use in Hermes
/memory add alice "Alice works at Acme Corp"
/memory retrieve --query Acme --limit 10
/memory retrieve --limit 10
/memory corroborate alice
/memory prune --max-nodes 500

Or import in your own agent:

from agentic_memory.hermes import setup

setup(agent=your_agent_instance)

Scoring System

Each memory entry gets a composite relevance score:

score = (3 × corroborations) + (1 × read_count) + (0.5 × recency_bonus)

• Corroborations increase when the same fact is reinforced (e.g., mentioned again by the user or confirmed by another source).
• Read count increases each time the memory is retrieved.
• Recency bonus decays over time — recently accessed memories score higher.

This ensures the most relevant, most frequently accessed, and most corroborated memories bubble up to the top during retrieval, keeping LLM context windows tight.

Auto-Pruning

Auto-prune runs as a background scheduler inside the MCP server process, controlled entirely via environment variables:

# Activate auto-prune: max 1000 nodes, prune when file > 5 MB, every 30 min
MEMORY_AUTO_PRUNE_ENABLED=1 \
MEMORY_AUTO_PRUNE_MAX_NODES=1000 \
MEMORY_AUTO_PRUNE_MAX_SIZE_MB=5 \
MEMORY_AUTO_PRUNE_INTERVAL_SECONDS=1800 \
  memory-mcp

Manual prune is also available via CLI or memory_prune tool.

Configuration

Development

# Install dev deps
uv sync --all-extras

# Run tests
uv run pytest tests/ -v

# Lint
uv run ruff check src/

License

MIT — Alexandre Rodenas