Graph memory loses to flat vector baseline on LongMemEval: F1 0.42 vs 0.47

CShorten30 · x · 2026-09-06

A controlled study testing whether graph memory beats flat retrieval for long-term agents found the opposite: researchers extracted each conversational turn into typed nodes and attributed edges, answered from two-hop subgraphs, and held the candidate-generation budget at five retrieval roots. On LongMemEval the graph scored token F1 0.42 vs 0.47 for the flat baseline, with a paired bootstrap over 500 questions showing a -0.050 gap (95% CI -0.085 to -0.016). The damage concentrated on questions requiring recall of a specific prior assistant turn, where judged correctness fell from 0.911 to 0.607—splitting turns into entities discards the surface form those questions need.

Original post →

More from coding & agent

coding & agent channel →