Keep evidence.
Bound authority.

A new study tests a boundary that token-budget metrics miss: whether an agent can reconstruct what the harness discards.

In a pre-registered synthetic experiment, truncation preserved success when a missing fact could be retrieved again but drove performance toward chance when it could not. A second-provider replication showed the same distinction. Separate ARC-AGI-3 ablations suggest summaries and notes can recover much of a harness gap without replaying earlier reasoning.

Evidence & limits. The ARC experiment covers one model, four games and one counted run per configuration/game. Component shares are fragile estimates, not a general ranking of memory strategies. Code, preregistration and analysis outputs are public; not all raw transcripts are included.

A PHREEQC geochemistry study finds that context-saving retrieval can fail after the agent has already computed the correct result.

Across 158 execution-verified questions, section-index retrieval reduced GPT-5.1 and GPT-5.2 accuracy by 11.3 and 7.9 percentage points; the tested Claude models showed no detectable loss. Replay traced failures to agents substituting their own higher-precision files for the simulator’s authoritative report and selecting a neighboring rounded answer. Explicitly identifying authoritative output improved the tested GPT models.

Evidence & limits. One scientific tool and question format, closed APIs, uneven repeats and model-specific decoding settings. These are author-reported experiments, not evidence that retrieval generally harms agents.

ScopeMem challenges the assumption that a trained memory prefix helps chiefly by remembering history. In its controlled GUI experiments, much of the benefit comes from revising the actor’s initial action.

Conventional cross-entropy-trained prefixes repaired mistakes but also broke 8–13% of previously correct decisions. ScopeMem conditions its prefix on the initial answer and trains preservation alongside correction, reducing the broken-correct share to 5.6% in the controlled ablation.

Evidence & limits. Offline step-level mobile-benchmark evaluation with one frozen Qwen3-VL-8B actor and recent history already visible. Online task completion remains untested. Two actor calls per step roughly double latency; fine-grained analyses use one training run.

VeriCUA trains a Judge to select completion checks and stopping points, while a frozen Probe inspects the live environment after the actor finishes.

Trained on OSWorld and tested on unseen mobile and web environments, the 4B system reports 74.08% macro-average reward agreement versus 68.72% for few-shot GPT-5.4, with 47.6–59.8% fewer Probe tokens. Targeted evidence collection lets a counterexample establish failure without inspecting everything.

Evidence & limits. Agreement is with benchmark executable evaluators, not proof of correctness. The reported 3.21-point downstream gain is Actor Pass@4, not single-attempt task success. Code is public, but the repository still says weights will be released. No independent reproduction here.

Microsoft announced Dynamic Permissions, Dynamic Agent Risk and Dynamic Conditional Access. Agents receive distinct identities, with authority scoped to an approved task and its duration, inside organization-wide policy.

The described responses are graduated: retain reads while removing writes, narrow reachable resources, require verified human approval, or isolate execution. The control boundary sits outside the model, rather than relying on the model to police itself.

Evidence & limits. An announced vendor architecture, not independently measured security effectiveness. The reviewed announcement does not establish general availability or deployment prerequisites for each component.

The UK Council for Science and Technology published advice calling for a single accountable owner and a map of AI capabilities to own, develop with partners, or access resiliently.

It recommends public-interest pilots comparing open-weight and domestic smaller models, national compute for trials, and consideration of a public-interest sovereign base model. The framework includes skills, data, institutions and the surrounding system, not only model weights or hosting location.

Evidence & limits. Newly published advice, not newly enacted policy, approved funding or a deployed model. The underlying letter is dated July 30, 2026, outside this edition’s event window; October 19 is its publication date.