The Evidence Was All There. The Retriever Never Reassembled It.
Relevant retrieval can still be incomplete retrieval.
A September 2026 study evaluated 2,000 span-grounded queries drawn from more than 1,200 agent execution trajectories. Under a fixed token budget, provenance-aligned retrieval units substantially improved complete-evidence retrieval compared with flat text windows, especially when the evidence required for one answer spanned multiple execution events.
The continuity failure
Fixed-window retrieval preserved relevant fragments while losing the transformation and event relationships needed to assemble a complete support set.
The GovKM interpretation
Lineage is not merely audit metadata. It can be retrieval structure. A consequential answer requires the evidence relationship, not simply the nearest fragments.
Continuity path: Source / execution events → Evidence / provenance units → Context / complete support set → Decision → Record → Institutional Memory → Future Reuse.
Source
Yiqi Wang et al., “When Does Execution Provenance Help Agent Memory Retrieval?” arXiv, September 22, 2026.



