Three Memory Layers, One Graph: A Technical Look at neo4j-labs/agent-memory — and Where Its Audit Trail Quietly Breaks

neo4j-labs/agent-memory gives agents three memory layers in one graph: conversation history, a long-term entity graph on the POLE+O model, and reasoning traces of what tools it called. Ships an MCP server and LangChain adapters. The audit trail links a reasoning step to the entities it touched, but only for tool calls mapped by hand; an unmapped call disappears from that query. 435 stars, Apache-2.0.
#KnowledgeGraphs #Neo4j #AgentMemory #GraphDatabase

neo4j-labs/agent-memory gives agents three memory layers in one graph: conversation history, a long-term entity graph on the POLE+O model, and reasoning traces of the tools it called. It ships an MCP server (16 tools by default) and integrations for LangChain, PydanticAI, CrewAI, Google ADK, and others. The audit trail links a reasoning step to the entities it touched via explicit :TOUCHED edges — but recording a tool call requires an explicit record_tool_call() call per invocation, so any tool call the developer doesn't instrument manually won't appear in that trace. 380 stars, Apache-2.0.