Graph-Based Agent Memory

Graph-based agent memory studies how AI agents can store and reuse information across long conversations, tool-use episodes, tasks, and changing environments. Systems commonly represent events, facts, skills, and their temporal or causal relationships in typed graphs or hypergraphs, then retrieve relevant subgraphs through traversal, ranking, or learned query-conditioned routing. Some also update, consolidate, revise, or forget memories using feedback, reinforcement learning, provenance, and validity checks; evaluations span conversational benchmarks such as LoCoMo and LongMemEval as well as tool-use and embodied-navigation tasks. Structured links can help with multi-hop reasoning and experience reuse, but graph construction and maintenance add costs, and results still depend on extraction quality, retrieval design, and fair comparisons with simpler vector or full-context methods.