Research Radar: Stored Lessons Make Agents Repeat Their Mistakes, a Decision-Model Controller Adds 19% Cost for No Gain, and Agents That Edit Their Own KV Cache Cut Cost 27.7%

This edition covers three preprints submitted between October 17 and 20, 2026 and listed on Hugging Face's Daily Papers on October 21 (the first two) and October 22 (Cache Engineering).

More Memory, Worse Agents (arXiv 2610.27406)

Many agents improve without weight updates by saving rules, summaries and skills inferred from past runs. The Indiana University authors identify a failure loop they call error reproduction. A wrong abstraction is retrieved, shapes the next execution, and that trajectory is then re-abstracted into the same error, "extending their persistence beyond direct reuse of the original faulty knowledge". Their fix, Anti-Persistence, keeps raw trajectories and outcomes but rebuilds task-specific guidance on demand and discards it after use. They report the highest held-out success and lowest average LLM-call cost among the memory designs they tested, across several benchmarks and backbones.

Good Judgment, Bad Advice (arXiv 2610.24649)

A pre-specified study on Terminal-Bench 2.1 kept a Claude Sonnet 5.5 agent loop fixed and varied only the controller consulted at decision points: none, rules, Claude Haiku 5.5, the typed decision model Jev 1.13, or a placebo sampling Jev's choice frequencies. Over 315 trials, "Jev raised cost per attempt by 19%" at an identical 73.0% pass rate, and cost 22% more than the placebo. Offline, Jev ranked finish requests best (AUROC 0.78 against 0.62 for rules) but was poorly calibrated. Its mean completion probability was 0.50 on runs that passed 87% of the time, so it kept asking for verification of correct work. "No controller ever chose to abandon a run."

Cache Engineering for Agents (arXiv 2610.28577)

Summarising or trimming history changes the prompt prefix, which invalidates cached state and forces costly re-prefill. Tsinghua's authors let the agent operate on its KV cache directly: drop excludes entries from future attention, and repos adjusts retained keys' position encodings. Because identical tokens can now map to different cache states, prefix matching is no longer enough, so they built AKV, a serving engine with a stateless API and stateful reuse inside. A rolling drop of old tool responses scored 1.7 points higher on average at 27.7% lower cost than threshold-triggered summarisation. AKV reached up to 1.2x output throughput and 1.5x faster P95 latency. Code is at thu-nics/AKV.

Analysis: Read together, the papers show agent add-ons failing in different ways. Memory can make a mistake permanent, a controller can be right on average yet wrong where it matters, and compaction can cost more than it saves. The practical test is the same for each: run the component against a placebo or a do-nothing baseline on your own tasks, measure both cost and pass rate, and check calibration, not only ranking. The cache paper matters for GPU budgets. If tool output dominates context, dropping it at the KV level while preserving the prefix cache is a cheaper lever than another summarisation prompt, though it needs a serving engine that tracks cache history.