I’ve released AgentMemory for .NET, an independent, native .NET implementation of persistent, graph-native memory for AI agents:
This became much more than a direct port. Several of the required pieces simply did not exist in the .NET ecosystem, so the Neo4j persistence and retrieval layers, extraction pipeline, dependency-injection architecture, Microsoft Agent Framework and Semantic Kernel adapters, MCP integration, and compatibility tooling were implemented natively in .NET.
As the neo4j-agent-memory, It supports three connected memory layers:
short-term conversation memory;
long-term entities, facts, preferences, and relationships;
reasoning traces, steps, and tool calls.
Retrieval includes vector, full-text, hybrid, and graph traversal. Compatibility is verified against the upstream Agent Memory TCK, with 178/178 Bronze, Silver, and Gold tests passing.
I’ve also opened a PR proposing a complete shopping-assistant sample for Microsoft Agent Framework, using the published AgentMemory packages:
I would love feedback from the Neo4j community—and contributors are very welcome. I’m particularly interested in discussing:
alignment with the wider Neo4j Agent Memory ecosystem;
areas where the .NET implementation could improve;
additional samples, integrations, and use cases;
and what the most useful long-term relationship with the Neo4j ecosystem might be.
@lyonwj, @michael.hunger and @zaid.zaim—you already know parts of the story, but perhaps not how far the implementation eventually went. I would really value your feedback on where it should go from here.
