The paper investigates why an LSTM agent outperforms a memory‑free MLP on a partially observable XLand‑MiniGrid task. By separating training seed, map, and evaluation randomness across 60 runs, the authors find that much of the 25‑percentage‑point gap is due to noise, and that providing the time step to a memory‑free MLP reduces the gap to a few points. A linear probe shows the LSTM’s state largely encodes the time step, indicating that the apparent memory advantage is largely an advantage of knowing episode time. "whyItMatters":"The study clarifies that the perceived memory benefit in this benchmark is largely attributable to time‑step awareness rather than true memory, guiding future baseline design and evaluation practices."
arXiv:2610.00604v1 Announce Type: cross Abstract: Vision-language-action policies often see only one or a few recent frames, which makes it difficult to evaluate how they use information that disappe...
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.
arXiv:2607. 20792v1 Announce Type: new Abstract: Memoir combines per-sample fast memory, shared slow parameters, variable-depth latent recurrence, and a future-latent energy objective.
arXiv:2606. 27472v1 Announce Type: cross Abstract: Large language model (LLM) agents operate over long, multi-session interactions in which facts change: a user moves, a price updates, a plan is revised.
The paper investigates how persistent memory in AI agents can lead to over‑trust in stale facts, creating a "Memory Trust Gap" that worsens as model capability increases. Using a benchmark with Benefit and Safety suites across Qwen3 models of varying sizes, the authors show that larger models are more prone to harmful over‑trust, especially when metadata is absent or misleading. They also demonstrate that mitigation strategies such as exposing metadata or pre‑resolving conflicts improve accuracy, but the effectiveness depends on model size and dataset.
arXiv:2606.24595v2 Announce Type: replace Abstract: Long-term memory promises LLM agents that grow more capable across sessions, maintaining an accurate, evolving understanding of the user that inter...
arXiv:2609.13922v1 Announce Type: new Abstract: Minibatch persistency reuses data instead of reading it: rather than drawing a fresh minibatch at every optimizer step, it takes K consecutive steps on...
arXiv:2610.27714v1 Announce Type: new Abstract: When does experience memory help a closed-loop LLM driving agent? Published gains cannot say, because the two factors we find to matter most, which war...
The study investigates how gate initialisation and recurrent regularisers (dropout and zoneout) affect learning long dependencies in one‑layer LSTM and GRU networks on synthetic tasks. By running 720 factorial experiments that cross conventional and chrono gate initialisation with up to six regulariser placements, the authors find that chrono initialisation consistently enables networks to reach validation criteria within 5000 updates, whereas conventional initialisation often fails or learns slowly. Dropout generally delays learning, zoneout only helps in limited conditions, and early input sensitivity provides limited predictive power for successful training.
arXiv:2607. 24667v1 Announce Type: new Abstract: A language model with a bounded working memory must repeatedly decide which stored items to keep.
arXiv:2606. 29178v1 Announce Type: new Abstract: When does retention matter for memory-augmented LLM agents?
The paper introduces a compact decoder that jointly handles gating and memory‑grounded hint generation for proactive assistants. Experiments on 28,058 decision points show that a single 1B‑parameter decoder can match or surpass two‑model pipelines, achieving higher precision and lower compute costs while preserving hint quality. Human evaluations confirm the benefit of memory grounding and demonstrate that the single‑model approach performs comparably to a larger 8B pipeline.