Execution-Provenance Units Beat Flat 512-Token Windows by 19 Points on Agent-Memory Retrieval

arXiv·medium signalXBlueskyLinkedInCopy link

The paper frames agent-memory retrieval as budgeted evidence completion and scores exact gold spans across 2,000 queries over 1,207 held-out ISETrace trajectories. Chunking memory into provenance units built from tool arguments and outputs raised Full Support@2048 by 19.07 points over flat 512-token windows, and 11.96 points above an oracle choosing the best of four flat chunk sizes. A zero-initialized residual R-GCN over typed provenance edges refines dense scores further. Builders storing agent traces should chunk by tool call, not by token count.