Graph memory underperforms flat retrieval on LongMemEval benchmarks

omarsar0 · x · 2026-09-02

A study evaluates whether graph memory outperforms flat retrieval for long-term agents. On LongMemEval, the graph approach achieved a Token F1 of 0.42, lagging behind the flat vector baseline's 0.47. Paired bootstrap testing confirmed a significant gap (95% CI -0.085 to -0.016). The performance hit is concentrated on questions requiring specific assistant turns, where accuracy dropped from 0.911 to 0.607 because splitting turns into entities discards necessary surface forms. However, the forgetting module performed well, successfully pruning 9.8% of nodes based on recency and access frequency.

Original post →

More from coding & agent

coding & agent channel →