Long-term agents, such as AI assistants that maintain memory across sessions, face a fundamental constraint: they must operate within a limited context budget. This means they cannot recall everything, only a subset of what is stored. A new paper on arXiv argues that the selection criterion should not be individual relevance alone, but collective sufficiency. The agent must recall evidence that, taken together, is enough to answer the current query.

Conventional memory access, the paper notes, retrieves and scores memories based on how relevant each one is to the query. The authors argue that this approach can miss what is needed. Even if each retrieved memory is relevant, the set as a whole may be insufficient. The paper's title, 'Recall What Is Missing,' points to the need to compose evidence from long-term memory in a way that fills the gaps.

Because this digest is based on a single source, there are no differing perspectives to compare. The abstract is brief and does not detail the proposed method, but the central claim is clear: under a limited context budget, memory recall should be evaluated at the level of the evidence set, not just the individual memory.