Mnemon is a memory agent that stores conversations as raw, dated records and uses a fast System 1 decision model (Jev) to quickly judge the relevance of records, while a slow System 2 LLM plans searches and composes answers. The agent consolidates records into topic timelines and value histories in the background, enabling efficient retrieval without rewriting conversations into structured formats. Experiments show Mnemon achieving high scores on LoCoMo and LongMemEval‑S with low context length and cost, and Jev outperforming LLMs in evidence separation and speed.
By Guangren Wang
The paper introduces LRE (Learned Relevance Eviction), a lightweight, CPU‑only, language‑model‑free scorer that learns which parts of an agent’s interaction history are task‑critical and preserves them verbatim. In experiments, LRE matches or surpasses baseline eviction policies on accuracy‑cost trade‑offs, recovers 93% of full‑history accuracy, reduces worst‑case prompt size by 52%, and outperforms dense and token‑pruning encoders in conversational memory while being 295–1569× smaller. The method also achieves superior budgeted answer quality on LoCoMo reading and can be trained annotation‑free, recovering 95% of supervised scorer performance.
By Nusrat Jahan Lia, Aritra Mazumder
The paper introduces Just-in-Time Memory (JitMem), a system that defers memory curation until a task is read, allowing a curator to synthesize task‑specific memory payloads based on the current query. Unlike traditional write‑time curation, JitMem retains raw trajectories and trains the curator using immediate task success, avoiding long‑horizon credit‑assignment issues. Experiments on ALFWorld, WebShop, and τ²‑bench show JitMem consistently outperforms both no‑memory agents and existing write‑time memory methods, with improvements of up to 16.3 absolute success‑rate points.
whyItMatters":"By curating memory at read time, JitMem enables more effective, task‑adaptive recall that directly improves agent performance across diverse benchmarks."
By Yefan Zhou, Yang Li, Zeyu Leo Liu, Semih Yavuz, Shafiq Joty
arXiv:2606. 12945v1 Announce Type: new Abstract: Long-running LLM agents accumulate interaction histories far larger than any context window, forcing a standing decision: what to encode deeply, what to forget, and what to retrieve under a fixed memory budget.
By Zhibao Chen, Qian Cheng
SkillGate is a method that trains agents to select the correct skill from a large slate during an episode by separating credit signals for skill selection and execution. It addresses the problem of selector credit starvation, where traditional outcome-rewarded RL fails to give sufficient credit to the skill-naming tokens, especially in long-horizon tasks. Experiments on five benchmarks show that SkillGate improves a 9B policy’s success rate from 40.8% to 53.2%, reduces exposure to misleading candidates, and requires fewer skill reads.
By Qingyao Li, Wenxiang Jiao, Shuai Shao, Kangning Zhang, Yuan Lu, Yi Guo, Weiwen Liu, Weinan Zhang, Yong Yu
arXiv:2606.22953v2 Announce Type: replace
Abstract: Context managers summarize or evict old messages so agents can operate beyond a finite window. This creates a reliability problem: removing history...
By Aman Mehta, Anupam Datta
The paper introduces Hindsight Structured Advice Distillation (HSAD), a method that trains a small advisor model to generate structured guidance for a frozen executor language‑model agent. By analyzing executor failures on a training set, a teacher writes advice that would have helped, and the advisor learns to produce this advice from input alone, adding minimal latency. Experiments on the FrontierFinance benchmark show that an 8B advisor boosts the qualification rate of Samaya’s research agent from 52.9% to 56.6%, surpassing the agent’s high‑effort configuration while cutting cost per query by 37%, and the approach also improves performance on WebShop, BIRD text‑to‑SQL, and ExpertLongBench.
By O. Ozan Koyluoglu
The paper investigates how a coding agent learns across a sequence of abstract reasoning tasks in the ARC-AGI-3 benchmark. The agent, operating on a frozen foundation model, writes and runs Python and shell scripts, retaining only its written artifacts as state. The authors develop a measurement protocol that traces each thought—any belief, rule, or plan committed to a file—through its lifecycle, revealing that most task-specific scripts are discarded at task boundaries and that the agent selectively forgets by rewriting knowledge into new scripts while accumulating unedited notes.
By Chen Wu, Josh Passenger, Yin Song
FM‑Bench is a new benchmark that tests large language model agents on long‑horizon decision making by having them run a football club for 20 in‑game years. The agent must manage a squad, trade players, negotiate contracts, invest in facilities and youth, set lineups, and respond to a board that can fire it, all using 26 tools and roughly 340–400 decision stops, with a deterministic engine producing a final score without human or LLM judges. The benchmark includes a solo track where each of 15 frontier models competes against a frozen scripted world, and an Arena track where the same models plus a scripted anchor share one 20‑year world, allowing the first head‑to‑head evaluation at this scale.
whyItMatters":"FM‑Bench provides a rigorous, large‑scale test of sustained, cumulative decision‑making in language‑model agents, revealing that managerial strategy—not computational scale or vendor—drives performance over long horizons."
By Tianyou Wang, Chongyang Gao, Kezhen Chen, Chen Dong, Yinghao He, Donghan Li, Wangcheng Xu, Hongjiu Zhang, Chi Li
LLM agents are increasingly evaluated on multi-week decision tasks in which the state that drives cost is never directly observed. On such tasks the final cost cannot say why an agent failed: it may have misread the world, or read it correctly and still failed to act (the knowing-doing gap).
The study investigates whether post‑training makes language models’ writing more predictable and whether this effect carries over when the same models act as judges. Using two 7‑billion‑parameter families (OLMo‑2 and Zephyr), the authors track changes from base models through supervised fine‑tuning (SFT) and direct preference optimization (DPO). They find that as writers, models produce increasingly familiar stories, but as judges, the models’ preference for predictable writing barely shifts, while they become more sensitive to story length, display order, and lose consistency on creative‑writing questions.
By Arman Nik Khah, Arvin Bahreini
FM‑Bench is a new benchmark that tests large language model agents on long‑horizon decision‑making by having them run a football club for 20 in‑game years. The agent must manage a squad, trade players, negotiate contracts, invest in facilities, set lineups, and respond to a board that can fire it, all while a deterministic engine aggregates the outcomes into a final score without human judgment. The benchmark evaluates six behavioral capabilities and compares 15 frontier models in solo and arena tracks, revealing that managerial behavior—not computational scale—drives performance.