AGM -- Agent Memory Management

agm-memory-mcp -- Provenance-tagged memory infrastructure for multi-agent AI systems, available as an MCP server.

Solves two problems current agent memory systems ignore:

1. Agent identity and trust -- knowing which agent wrote what and whether it can be trusted
2. Memory poisoning -- detecting and containing malicious or corrupted memories before they propagate through agent reasoning

Existing systems (MemGPT, Mem0, HippoRAG) have no memory states, no trust provenance, and no causal ordering. AGM is infrastructure for pipelines that need those properties.

Install

pip install agm-memory-mcp

Python import name:

import agent_memory_mcp

CLI command:

agm-memory-mcp --help

Requirements

• Python 3.11+
• A running AgentMemory HTTP API endpoint

This package is the MCP bridge. It does not embed the full backend. To run the backend locally (needs only Docker):

git clone https://github.com/Deep-De-coder/AGM
cd AGM/agent-memory
docker compose up --build        # API on http://localhost:8000

Configuration

Set either CLI args or environment variables.

Environment variables:

• AGENT_MEMORY_API_URL (default: http://localhost:8000)
• AGENT_MEMORY_API_PREFIX (default: empty string)

Examples:

# Linux/macOS
export AGENT_MEMORY_API_URL=https://your-api.example.com
export AGENT_MEMORY_API_PREFIX=/api/v1
agm-memory-mcp

# Windows PowerShell
$env:AGENT_MEMORY_API_URL="https://your-api.example.com"
$env:AGENT_MEMORY_API_PREFIX="/api/v1"
agm-memory-mcp

Or pass options directly:

agm-memory-mcp --api-url https://your-api.example.com --api-prefix /api/v1

Health Check

Validate connectivity before starting MCP transport:

agm-memory-mcp --check

Shows resolved URL and API status. Exits 0 on success, 1 on failure.

Run as Module

python -m agent_memory_mcp

Equivalent to running agm-memory-mcp.

MCP Client Config Example (Claude Desktop / compatible)

{
  "mcpServers": {
    "agm-memory": {
      "command": "agm-memory-mcp",
      "args": [
        "--api-url",
        "https://your-api.example.com"
      ]
    }
  }
}

You can also use env vars instead of args.

Included MCP Tools (14)

1. write_memory
2. read_memory
3. query_memories
4. get_safe_memories
5. get_trust_score
6. get_provenance
7. flag_memory
8. register_agent
9. check_violations
10. acknowledge_violation
11. get_notifications
12. run_rules_check
13. consolidate_memories
14. get_rules_reference

Rule Model

AGM includes 15 embedded detection rules (RULE_001 through RULE_015) for:

• write flooding
• trust-chain contamination
• contradiction detection
• rapid rewrite anomalies
• high-stakes unverified content
• inter-agent session violations
• expired safety context
• anonymous agent writes
• bulk near-duplicate content
• trust cliff detection
• behavioral drift
• causal orphan detection
• anergy bypass attempts
• high-taint writes (causal contamination)
• sandbagging (evaluation vs. production behavior gap)

Use get_rules_reference to inspect rule metadata at runtime.

License

MIT