Remembering the right context and persisting through a difficult task are complementary capabilities. They can also amplify each other’s mistakes. An agent that retrieves a wrong preference and diligently follows it may be less useful than one that forgets and asks.
Meta’s October 12 Muse Wonder and Muse Wonder Full release makes this interaction worth evaluating directly. The new question extends beyond our Muse stack analysis and Small Business assessment: when memory improves a long task, how much comes from recall, how much from sustained execution, and how well does the combination recover from a changed premise?
What Meta announced
The release makes Wonder available through the Meta Model API and Muse Code, with Wonder Full available to Muse Code subscribers. The announcement describes a subsequent subscriber-first rollout into Muse; it should not be read as proof that every personal-agent session already uses Wonder. October 12 announcement, updated October 13
Meta calls relevant context carried across conversations Active Memory. Active Focus describes continued work through blockers and new information, including clarification. The research post connects these ideas to MemoryBench, FocusBench, and SWE-Marathon, and argues that stronger base models improve generalization beyond post-training targets. These are publisher descriptions and interpretations, not findings reproduced here. Research account
The mechanism is broader than a long context window. My interpretation is that a useful continuing agent needs a selection policy for past information and a control policy for continuing work. A large window gives those policies capacity; it does not decide what to retain or when to reconsider.
Two capabilities, four experimental conditions
The first study I would run has four conditions: memory absent with a short execution budget; memory present with that budget; memory absent with an extended budget; and memory present with the extended budget. Keep the selected model, tools, task inputs, and evaluator constant wherever the product permits.
This separates a memory effect from the value of additional attempts. It also exposes interaction. If the combination beats the sum of the isolated gains, memory may be helping the agent resume useful work. If the combination performs worse after a correction, persistence may be reinforcing stale context.
Treat this as a proposed application-level ablation. The public release does not establish that customers can disable internal Active Memory or Active Focus independently. Where those controls are unavailable, approximate the intervention with external history access and fixed execution budgets, label the approximation, and avoid claiming to isolate internal training mechanisms.
Use repeatable tasks with several stages: an initial preference, a delayed request that needs it, a correction, and a later request where the old fact should no longer apply. Include a control where the earlier detail is irrelevant. A memory system that always retrieves something can increase confident errors.
Evaluate the decision to remember
A recall question asks whether a fact can be recovered. A working agent also needs to decide whether the fact applies to the current task.
Consider an illustrative assistant helping a user plan an event. A previous conversation says they prefer outdoor venues. A later conversation says the current event is for a relative with a different accessibility requirement. The earlier preference is true, but it may be irrelevant or subordinate to the new constraint.
The evaluation should therefore score retrieval precision, correct attribution, temporal validity, and appropriate abstention. Preserve the source and date of every target fact in the test fixture. Ask reviewers whether the final plan used the right person’s preference for the right event, rather than merely repeating a remembered phrase.
Also test conflicting and low-confidence material. A model-generated summary should not gain the authority of a direct user statement merely because it persisted longer. An effective agent needs to distinguish a recorded observation from a current instruction.
These are design criteria for an application, not observed Wonder defects. The study would need synthetic or appropriately consented information and an explicit retention policy before collecting traces.
Persistence needs a changing-goal test
Continuing through a temporary failure can help. Continuing through an invalidated objective wastes time and may produce the wrong artifact. A task can change because the user corrects a requirement, a source is withdrawn, or a necessary dependency proves unavailable.
I would insert a controlled change halfway through a long task and inspect the resulting work. Does the agent update its plan, revisit affected assumptions, and explain which completed work can be retained? Or does it preserve the original trajectory and attach the new requirement only to its final explanation?
Measure useful progress after the change, discarded work, clarification quality, and time to a coherent revised artifact. Include a legitimate stop condition. Rewarding continuation on every case would train the evaluation to prefer stubbornness.
The economic denominator should include the entire attempt, including work abandoned after a correction. More internal steps can reflect productive exploration or difficulty escaping a mistaken plan. Step count by itself cannot distinguish them.
Read each benchmark row as a separate experiment
The model card says most Wonder results use maximum reasoning effort. It distinguishes shared-harness internal coding comparisons from externally reported product/harness results. Its SWE-Marathon row mixes agent products and uses xhigh for Wonder Full. OSWorld’s primary metric is partial reward, with binary completion secondary. Model-card methodology
MemoryBench is described as an internal 64K–100K conversation evaluation with distractors and updates, scored by an LLM judge. FocusBench is also internal. These can inform hypotheses without establishing cross-device memory lifecycle behavior or performance on an arbitrary customer’s history. Memory and focus evaluation definitions
My implication for a purchasing team is to preserve the metric and harness with every score. Do not average strict completion and partial credit into an unlabeled success rate. Do not infer that a long software project score isolates memory quality when planning, tool use, and extra compute also contribute.
A useful local comparison should report complete-task acceptance alongside intermediate credit. The latter helps diagnose progress; the former reflects whether the user received a usable result.
What independent evidence should add
I would prioritize temporally separated tasks that the evaluator and development team freeze before tuning. Vary names, document layouts, and tool order to reduce near-duplicate transfer. Blind artifact reviewers to the model and condition. Preserve failures and missing results.
For automated judging, audit a stratified sample with domain experts, especially disagreements and apparent high-scoring failures. A judge can reward a coherent narrative that silently uses the wrong person or period. Publish disagreement rates and examples of corrected grades.
For business value, track how much reconstruction the user avoids and how much correction the user performs. A remembered fact saves time only if locating, checking, and correcting it costs less than re-entering the relevant context. Include memory maintenance in the operating cost.
Current action permissions remain separate from remembered preferences. A retained instruction does not renew an expired mandate; long-running and resumed work still needs current data access, applicable approval, and shared budget accounting. This is an application obligation, not a claim that Meta implements the ContextOS contract.
The release makes a promising combined capability available for investigation. My adoption criterion would be better accepted work after interruptions and corrections, with measured savings in user effort. That would demonstrate useful continuity, rather than simply a stronger ability to recall and continue.
Dated source ledger
Sources opened October 19, 2026:
- Muse Wonder launch: October 12, updated October 13; product access and rollout distinctions.
- Muse Wonder research post: October 12; Active Memory/Focus framing and generalization argument.
- Model card: October 12 release identified in the card; evaluation definitions and configuration differences. No card result was rerun.