Stop resetting your agents' brains after every chat

What it does Memary wraps a persistent memory system around LLM agents. It stores agent interactions and extracted entities in a graph database (Neo4j or FalkorDB), then retrieves context through recursive, multi-hop subgraph searches rather than dumping the whole knowledge base into the prompt. The current ChatAgent class bundles memory modules, a routing agent, and the knowledge graph together into one package.

The interesting bit The memory module takes its cues from Microsoft Research’s K-LaMP paper, pairing a time-stamped Memory Stream that logs the breadth of encountered concepts with an Entity Knowledge Store. When a query arrives, the system carves out a local subgraph of relevant entities up to two hops away, and only calls out to external search if the graph draws a blank. Agent responses are written back to the knowledge graph, so the system theoretically grows smarter as it runs.

Key highlights

• Human-memory-inspired architecture with auto-updating Memory Stream and Entity Knowledge Store
• Recursive retrieval and multi-hop reasoning over Neo4j or FalkorDB to avoid full-graph scans
• Supports local models via Ollama or OpenAI APIs, with optional vision, location, and stock tools
• Multi-graph support in FalkorDB for isolating memory contexts across different agents
• Custom tools can be injected into the built-in LlamaIndex ReAct agent interface

Caveats

• The bundled ReAct agent is labeled demo-grade and scheduled for removal; provider-agnostic support is still pending
• “Rewind memories” and full decoupling from the agent implementation are marked “coming soon”
• Running it requires API keys for OpenAI, Perplexity, Google Maps, Alpha Vantage, plus graph database credentials, and Python must be <= 3.11.9

Verdict Worth exploring if you are building long-running agents that need to accumulate user context without paying for infinite context windows. Skip it if you want a universal, drop-in memory backend today; the current ChatAgent coupling and bundled agent make it closer to an opinionated starter kit than a pluggable layer.