Neo4j Agent Memory
The best open-source repository for building a unified memory layer for AI agents via knowledge graphs. Models short, long, and reasoning memory with a clean ontology and extraction algorithms. Source: User, 2026-05-19
Three Memory Types
| Type | What it stores | Implementation |
|---|---|---|
| Short-term | Conversations, messages, per-session history | Vector and text search over recent interactions |
| Long-term | Entities, preferences, facts | Knowledge graph using POLE+O model (Person, Object, Location, Event, Organization) |
| Reasoning | Reasoning traces, tool usage, similar task retrieval | Helps agents learn from past decisions and avoid repeated mistakes |
Architecture
Multi-stage entity extraction pipeline: spaCy for NER, GLiNER for domain-specific entities, LLM for complex extraction. Relationship extraction via GLiREL. Background enrichment from Wikipedia and Diffbot. Geospatial queries. MCP server with 16 tools. Source: GitHub README, 2026-05-19
The reviewed ontology diagram makes the graph model concrete: user, fact, preference, document/chunk, POLE+O nouns, and custom objects are linked through relationships such as PART_OF, NEXT, and SUPERSEDED_BY. This is the practical reason Neo4j Agent Memory is not just "vector memory": it gives agents a typed relationship layer for entities, evidence, document chunks, preferences, and reasoning traces. Source: X/@pauliusztin_ and local image review, 2026-07-03
Current README features include hosted NAMS, a self-hosted Bolt path, shared Python and TypeScript SDKs, a create-context-graph app scaffold, bring-your-own model configuration, buffered writes, graph adoption, consolidation primitives, an eval harness, and explicit :TOUCHED audit edges from reasoning steps to entities. Source: GitHub README, checked 2026-07-03
Framework Integrations
LangChain, Pydantic AI, Google ADK, Strands, CrewAI, LlamaIndex, OpenAI Agents, and Microsoft Agent. Apache 2.0 licensed. Source: GitHub README, checked 2026-07-03
Status
Created January 2026 by Neo4j Labs. As of the 2026-07-03 source check: 355 stars, 82 forks, 78 open issues, default branch main at f29ae8dd86221bd5ace94d6c39711c38b0a6cdde, latest public tags v0.4.0, python-v0.5.0, and typescript-v0.4.0. The project remains experimental and community-supported. Source: GitHub API and git ls-remote, checked 2026-07-03
Relevance to Kevin's Wiki
Kevin's wiki implements a similar three-layer memory architecture but with flat markdown files instead of a graph database: short-term (chat context), long-term (wiki pages with compiled truth + timeline), reasoning (skills, patterns, concepts). The POLE+O ontology maps to Kevin's wiki categories: people, projects (objects), places, decisions (events), companies (organizations). Neo4j agent-memory offers a complementary approach for agent systems that need real-time graph queries rather than file-based search. Source: compiled from wiki, 2026-05-19
Timeline
- 2026-07-03 | Deep-reviewed the local ontology image and refreshed the current source snapshot: 355 stars, 82 forks, 78 open issues,
mainHEADf29ae8d, Python0.5.0, TypeScript0.4.0, hosted NAMS, shared SDK model, eval harness, buffered writes, and graph-adoption features. Source: X/@pauliusztin_, 2026-05-18; Source: GitHub README/API, checked 2026-07-03 - 2026-05-19 | Page created from User recommendation and GitHub research. DecodingAI article forthcoming. Source: User + GitHub, 2026-05-19
- 2026-05-18 | Paul Iusztin called it the best open-source repo for a unified agent memory layer via knowledge graphs (short/long/reasoning model, ontology, extraction). 477 likes, 636 bookmarks. Source: X/@pauliusztin_, 2026-05-18