Agent memory is moving from similar to useful
One system lifts end-to-end success from 16.0% to 76.0% by scoring memory on utility. Another improves accuracy by 10.9 points while cutting cost. The catch: confidence-based retrieval only pays off when budgets are tiny.
- In a preprint, Jetriever achieves 76.0% end-to-end application success on InMind tasks, compared to 16.0% for existing systems, and full-pool utility scoring costs approximately $0.03 per query. Self-reported m:…8130
- In a preprint, the ARISE architecture improves agent accuracy on SOP-Bench by 10.9 pp, and ARISE-Teacher reduces per-task costs by 1.0% compared to baseline. Self-reported m:…9770
- In a preprint, confidence-based retrieval outperforms fixed-depth policies only when the average retrieval budget is extremely scarce (0.5 passages per query), and confidence-based gating for depth selection performs no better than random chance. Self-reported m:…8141
- In a preprint, WPET allows for the systematic determination of evidence admissibility following context changes, and agent skill reliability is compromised by changes in permissions, dependencies, and erroneous specifications. Self-reported m:…0966
What changed
Most agent memory works by similarity: find the past note that looks most like the current question. A preprint on a system called Jetriever argues that is the wrong score. Utility-based retrieval using Jev outperforms similarity-based retrieval on indirect-query tasks, where the useful memory does not look like the question at all. The gap is large: Jetriever achieves 76.0% end-to-end application success on InMind tasks, compared to 16.0% for existing systems. Full-pool utility scoring is cost-effective for agent memory systems at approximately $0.03 per query, so scoring everything on usefulness is not a luxury purchase. It is, on these numbers, the difference between a memory that works and one that mostly retrieves plausible-looking irrelevance. m:…8130
A second preprint, ARISE, attacks the same problem from the learning side. The ARISE architecture improves agent accuracy on SOP-Bench by 10.9 pp, and ARISE-Teacher reduces per-task costs by 1.0% compared to baseline. ARISE-Teacher outperforms GEPA on tau^2-bench by 4.5 pp. The system works by self-evolving through iterative failure attribution and rule consolidation, which in plain terms means the agent looks at where it went wrong, writes down a rule, and gets better at the routine procedure next time. Accuracy up and cost down at the same time is a rare combination, and it suggests that memory is not just storage. It is the record of what the agent has learned not to do again. m:…9770
When smart retrieval does not pay
A third preprint is the useful sceptic in the set. Confidence-based retrieval outperforms fixed-depth policies only when the average retrieval budget is extremely scarce, at 0.5 passages per query. At more generous budgets, the clever approach stops winning. Confidence-based gating for depth selection, choosing between 1 and 5 passages, performs no better than random chance. And the latency overhead of the confidence probe often negates performance gains unless the gate skips retrieval for the majority of queries. The lesson is not that adaptive retrieval is useless, but that it is a tool for scarcity. If you can afford to fetch a fixed set of passages, fetch them, and spend the engineering time elsewhere. m:…8141
The fourth preprint, SCAFFOLD, asks what happens when a memory or a skill outlives the conditions it was validated under. WRCs, or Warranted Reuse Claims, serve as the appropriate unit of prospective trust for versioned agent skills, which is a formal way of saying that trust should attach to a specific version of a skill with specific evidence behind it. WPET, or Warrant-Preserving Evidence Transport, allows for the systematic determination of evidence admissibility following context changes. And agent skill reliability is compromised by changes in permissions, dependencies, and erroneous specifications between validation and invocation. A skill that was safe last month may not be safe today if its permissions changed, and SCAFFOLD is a framework for noticing that before the agent acts on stale trust. m:…0966
What this means for builders
Put together, these papers describe a memory layer growing up. The first step is to score memory on whether it helps the task, not whether it resembles the question, because Jetriever's 76.0% against 16.0% is too large to ignore. The second step is to let the agent consolidate its failures into rules, as ARISE does, so memory compounds instead of just accumulating. The third step is the one most teams skip: track the warrant. When a skill is reused in a new context, ask whether its evidence still applies, because a permission change or a dependency change can silently void the original validation. A memory system that cannot answer that question is a liability dressed as a feature, however good its retrieval scores look. m:…8130 m:…0966
What is still unresolved
- Whether Jetriever's utility advantage holds on enterprise tasks beyond the InMind benchmark.
- How ARISE's rule consolidation behaves over long deployments with many accumulated rules.
- Whether teams will adopt formal warrant tracking, or keep reusing skills on trust.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.