Your Agent Gets Dumber the Longer It Runs. Here’s How a Graph Fixes It.

At Agent Memory Night in San Francisco, FalkorDB’s Liz Zhang argued that long-running AI agents need memory systems designed as fast, structured “SSDs” rather than relying on full context histories or flat vector search alone. She described working, episodic, semantic, and procedural memory, proposing a two-loop architecture in which background processes extract and update facts while each agent turn retrieves only relevant, current information. The presentation contends that graph-based memory can improve accuracy, reduce latency and token costs, and prevent cross-user data leakage by modeling relationships, closing outdated facts instead of appending conflicting information, and assigning each user an isolated graph. Citing published evaluations and the open-source agent-memory-bench, Zhang reported that graph retrieval outperformed vector approaches on several reasoning tasks and that FalkorDB delivered low-latency graph, vector, and full-text queries under concurrent workloads, while acknowledging that graph memory does not solve poor information extraction and may be less suitable for questions requiring exact raw-text quotations.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.