When a language agent's execution history grows beyond its context window, its memory system must decide what to retrieve under a hard token budget. The problem is that supporting evidence may be scattered across multiple execution events, and conventional retrieval methods can miss the full picture. A new arXiv paper proposes EPGM, a system that uses execution provenance—the record of how events are causally linked—to guide retrieval.
EPGM is designed to reconstruct complete evidence chains rather than returning isolated snippets. By modeling the provenance of each execution step, the system can identify which events are jointly necessary to support a given query, then allocate its limited token budget across those events. This contrasts with approaches that treat each event independently, which risk fragmenting the evidence.
The paper's abstract emphasizes the challenge of retrieving complete supporting evidence under a hard token budget, and positions EPGM as a response to that challenge. While the full details of the method are not in the abstract, the core idea is clear: memory retrieval for agents should be budget-aware and provenance-driven, not just relevance-ranked. The work speaks to a growing need for efficient, reliable memory in long-horizon agent tasks.