Paper: knowledge-graph memory with forgetting beats flat vector retrieval for agents

burkov · x · 2026-09-07

A new paper asks whether representing conversational history as a knowledge graph improves retrieval quality for long-term LLM agents versus flat vector retrieval, and whether an explicit forgetting mechanism can keep the graph from growing without hurting performance. The motivation: agent memory stores balloon over months of interaction, raising costs and noise while flat retrieval struggles with relational and multi-hop questions.

The pipeline extracts typed nodes and attributed edges per dialogue turn, retrieves a two-hop subgraph around the top-5 matching entities, and periodically prunes nodes whose importance score (recency, access frequency, degree, age) falls below a threshold, evaluated on a 500-question benchmark.

Original post →

More from coding & agent

coding & agent channel →