Multi-Agent Shared Graph Memory: Building Collective Knowledge for Agents
summary
Argues that multi-agent systems fail from lack of shared memory and proposes a concrete architecture using Neo4j + Quadrant where Neo4j is the persistent world model and Quadrant stores long text. Demonstrates a four-agent example (researcher, planner, executor, critic) with runs that hydrate once, reason, act, and write back. Preserves provenance and disagreements via a contradicts edge instead of overrides, warns of schema sprawl, stale nodes, and urges one-query-per-run hydration.
AI-written summary · about
chapters (12)
- 0:00Introduction to talk
- 0:29Memory problem in agentic systems
- 2:37Root cause of memory issues
- 4:23Why use graphs for memory
- 5:46Three-layer architecture explained
- 9:24Example with four agents
- 13:04Graph schema as ontology
- 14:30Handling conflicts in graph
- 15:59Full agent loop breakdown
- 17:05Cumulative reasoning benefits
- 17:27Common pitfalls and fixes
- 19:14Conclusion and key takeaways
related talks
15mBuilding Evolving AI Agents Via Dynamic Memory Representations Using Temporal Knowledge Graphs
18mArchitecting Agent Memory: Principles, Patterns, and Best Practices
28mAgentic Memory
26mStore Your AI Agent's Memory and Context in a Knowledge Graph
29mTransparent Multi-Agent AI
01h 49mBuilding Memory Augmented Agents with MemoRizz
31mIs My Brain a Knowledge Graph? A Human Take on Connecting the Dots
58mAdvanced Autonomous AI Agents: Memory, Collaboration and Evolution in Action