Using Graphs to Fine-Tune Agents
summary
Graphs store every agent run as an execution graph in Neo4j and RAG retrieves prior successful runs as context to decide next steps, not just documents. Workflows become DAGs where each node is an LM call or function call and evaluation is backpropagated to optimize prompts and code by backpropagate to parameter nodes. An open source memory repo with ~1,500 stars demonstrates entity tracking and personalization. The system records successes and failures, learns probabilistic communication channels between agents, and scales from single-agent tuning to multiagency topologies.
AI-written summary · about
chapters (6)
- 0:00Introduction and inspiration
- 1:48Minsky and memory theory
- 4:26The Memary project
- 6:06Computational workflow graphs
- 8:13Optimization phases
- 10:35Future goals and conclusion
related talks
29mTransparent Multi-Agent AI
51mMemory Masterclass: Make Your AI Agents Remember What They Do!
17mContext Graphs for Explainable, Decision-Aware AI Agents
30mAgent Reasoning with Graph World Models
30mLive-Coding Graph-Native Agents: Neo4j AuraDB + Microsoft Agent Framework in Action
20mWhy your agents need decision traces, not just documents
29mThe AI Agent Memory Landscape
31mBuild Intelligent AI Agents With Context Engineering