Most agent memory systems in production are scored on recall. You log a conversation, extract facts, embed them, and at query time pull top-k by similarity. The eval harness asks: did the right memory come back? Three arXiv papers from the week of October 5–12 all argue, from different angles, that this is the wrong question — and the fixes land in different parts of the stack.
Relevance Is Not Sufficiency (Oct 7) makes the set-level argument: the top-ranked memories by relevance can each be individually on-topic and still, collectively, omit a complementary fact the answer depends on. That's a different failure than retrieving noise. Your ranker is working; the retrieved set just doesn't close the evidence gap. Any system tuned on recall@k will happily report success here, because each item it returned was relevant.
The Right Memory in the Wrong Context (Oct 5) attacks it from the governance side: a memory can be relevant and still inadmissible — it belongs to a different user, violates a policy, or sits in an incompatible lifecycle state (superseded, deleted, scoped to a session that ended). Final-answer accuracy doesn't catch this. If the agent answers correctly using a fact it shouldn't have been allowed to see, your metrics say you passed. For anyone running multi-tenant memory, that's the whole ballgame.
RECAST (Oct 7) is the retrieval-side response to the same problem: adaptive routing across multiple sources for the case where no single source contains the evidence, rather than the retrieval-centric agentic RAG most systems ship today.
What this changes architecturally
Two concrete implications. First, admissibility belongs in the retrieval predicate, not in a post-hoc filter or a system prompt. Tenant, policy state, and fact lifecycle are query-time constraints on the index. If you're doing ANN search and then filtering results, you silently lose recall on the admissible set — and you've already paid to embed and rank things the user can't see.
Second, sufficiency needs set-level selection or an iterative gap-check, not a bigger k. Raising k to improve coverage mostly adds near-duplicates, which is precisely the structural point a related paper, *Beyond RAG for Agent Memory*, makes: agent memory is a bounded, coherent stream with highly correlated spans, unlike the heterogeneous corpora RAG was designed for. Fixed top-k returns redundancy, and generic prompt pruners — built assuming retrieved passages are diverse and noise is separable — can delete temporally linked prerequisites.
Separately in the same week, a placebo-controlled study put holes in the structure-aware chunking assumption. Worth reading with the above: both are reminders that this field's defaults were inherited, not validated.
If you own a memory layer, the cheap action is adding negative controls to your eval — cross-tenant memories, superseded facts, questions requiring two complementary memories — and seeing what your current recall numbers hide.