Open-source memory layer for autonomous agents built around a `ChatAgent` class that wires three components: a ReAct-based routing agent, a knowledge graph (Neo4j or FalkorDB) for entity storage, and dual memory modules (Memory Stream for breadth, Entity Knowledge Store for depth). Every agent response is parsed into typed entities and written back to the graph; at query time a recursive subgraph retrieval extracts relevant context and injects it alongside ranked top-entities and a condensed chat history into the LLM prompt. A Streamlit dashboard lets developers inspect and replay the evolving memory state.
- Storage
- Knowledge graph — Neo4j (primary; free Aura cloud tier supported) or FalkorDB (added v0.1.4; multi-graph mode enables isolated memory spaces per agent in multi-agent setups). Entity nodes carry timestamps; a parallel in-process Entity Knowledge Store tracks per-entity frequency and recency scores. LlamaIndex is used for document-to-node ingestion into the graph store.
- Retrieval
- Recursive subgraph retrieval: extracts key entities from the incoming query, builds a bounded subgraph (max depth 2) per entity, then joins multiple subgraphs via multi-hop reasoning when several entities are present. Context window is assembled from the retrieved subgraph, the top-ranked entities by frequency+recency, and an LLM-summarized chat history. Falls back to an external Perplexity search (mistral-7b-instruct) when no relevant nodes exist in the graph.
- Self-host
- Self-host: moderate
- License
- MIT
- Pricing
- Open source (MIT), free to self-host; Neo4j Aura free tier covers the graph backend · Free / OSS
- GitHub stars
- 2,636
- Last release
- 2024-10-22
- Last commit
- 2024-10-18
- First catalogued
- 2026-06-28
Strengths
- Inspectable knowledge-graph memory: agent actions are viewable and queryable via Neo4j or FalkorDB consoles
- Recursive subgraph retrieval with bounded depth (max 2) reduces latency vs full-graph scans
- Dual memory model separates breadth (Memory Stream) from depth (Entity Knowledge Store) for nuanced context ranking
- FalkorDB multi-graph support enables isolated per-agent memory spaces in multi-agent setups (added v0.1.4)
- Streamlit dashboard for in-browser memory inspection and timeline replay; permissive MIT license
Watch out
- Deprecated/abandoned: no repo push since Oct 2024 (~20 months as of 2026-06-29); PyPI package stuck at v0.1.3 (May 2024), behind the v0.1.5 GitHub tag; 14 open issues with no maintainer responses
- Python version cap (<=3.11.9) blocks use in Python 3.12+ environments without patching
- Requires a running graph database (Neo4j or FalkorDB) — adds operational overhead vs in-process memory stores
Best for
- Research and prototype agents that need an inspectable, queryable knowledge-graph audit trail of their actions
- Multi-agent systems requiring isolated per-agent memory spaces via FalkorDB multi-graph, where a Neo4j/FalkorDB backend is already available
- Teams prepared to fork and maintain the library, given upstream is inactive since late 2024
How it integrates
Benchmark results
No sourced results yet.
Sources
- Monthly card re-verification — GitHub API repo metadata: 2,637 stars, license MIT confirmed, archived=false, last commit on the default branch 2024-10-18 (third-party)
- https://github.com/kingjulio8238/Memary (vendor)
- https://kingjulio8238.github.io/memarydocs/concepts/ (vendor)
- https://pypi.org/project/memary/ (vendor)
- GitHub API re-check — 2635 stars, MIT license confirmed, not archived, still last pushed 2024-10-22 (~21 months dormant as of 2026-07-23); PyPI still stuck at v0.1.3 (May 2024). Deprecated verdict confirmed, unchanged (third-party)
Last verified 2026-08-01 · updated by refresh-framework-cards