When Does Execution Provenance Help Agent Memory Retrieval?
4.00T1 sourcearXiv cs.MA
Source record
Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2609.25913.
Pipeline notes
The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.
SummaryA research paper reframing agent memory retrieval as budgeted evidence completion, using source-aligned provenance units from tool arguments/outputs and a zero-initialized R-GCN to refine dense retrieval scores. Tested on 2,000 queries over 1,207 trajectories, it reports a 19.07-point Full Support@2048 gain over 512-token flat chunks.
Why it mattersPinpoints a concrete failure mode in flat-chunk agent memory and reports measured gains from typed graph propagation. Relevant to anyone designing retrieval layers for long-horizon agents.
Cited by
No citations on record.