StackMap
Subscribe

agent-memory vs metronix-memory

Neo4j Labs' graph-native agent memory: conversations, a POLE+O entity knowledge graph and reasoning traces in one store, with a 16-tool MCP server and hosted or self-hosted backends. — versus — Self-hosted agent memory stack in Docker: Postgres + Qdrant + Neo4j hybrid retrieval, a temporal knowledge graph, ontology layer and freshness checks behind one MCP-native API.

The curated verdict

Both put agent memory on Neo4j and expose it over MCP. Neo4j Labs' version is graph-purist (POLE+O entities, reasoning traces); Metronix wraps the graph in a hybrid retrieval service with vector and sparse tiers alongside it.

agent-memorymetronix-memory
Stars51395
Forks979
LanguagePythonPython
LicenseApache-2.0Apache-2.0
Last activity7 days ago3 days ago
Topicsmemory, knowledge-graphsmemory, rag, knowledge-graphs
Curated connections76

agent-memory — the curator's take

The pick when memory has to be queryable as a graph instead of a black box: entities resolve and dedupe, reasoning steps get explicit :TOUCHED audit edges to the entities they used, and you can adopt an existing Neo4j graph as long-term memory rather than re-ingesting. Multi-tenant scoping, buffered writes, consolidation primitives and an eval harness are already in the box, and the hosted NAMS tier lets you start with no database to run. Caveats: Neo4j Labs marks it Experimental and community-supported; extraction stacks spaCy/GLiNER/GLiREL plus an LLM pass, so ingest costs real time and tokens; and if you don't want a graph database in the stack at all, a Postgres- or file-backed layer is far less machinery.

metronix-memory — the curator's take

Reach for it when you want one self-hosted box doing dense + sparse + graph retrieval and keeping long-lived facts fresh, and you're willing to run four datastores to get it. Skip it if you only need per-project session recall — a SQLite memory plugin is a fraction of the operational surface. Read the benchmark table with care: they are the author's own N=1 runs under a self-defined protocol, and the honest signal in them is that retrieval scores (Recall@10 85-95%) far outrun end-to-end answers (53-63%) — finding the evidence was never the hard part.