Muse Raises Memory’s Stakes

If You Only Read One Thing

Remembering a correction can make the next assignment worse. Meta’s Muse Wonder improves sustained agent work, while Katie Parrott’s context rebuild exposes what happens when yesterday’s feedback becomes permanent law. These developments raise a question beyond memory capacity: which past information deserves authority over the current task? Better recall increases the value of making that distinction before an agent starts.

Muse Earns a Place in the Comparison

Meta has a credible new agent model, but evaluating it requires choosing between two different products. Muse Wonder can enter an existing API-based workflow. Wonder Full requires Muse Code, so its strongest results cannot simply be assigned to the model a custom agent will call.

The October 12 release is a substantial improvement over Spark. Artificial Analysis’s independent evaluation puts Wonder and Full at 52 and 56 on its broad intelligence index, versus Spark 1.3 at 48. On its executable terminal tasks, Spark’s 33% rises to 52% and 57%. Those results establish meaningful progress without establishing superiority on every coding workload.

The more interesting evaluation question concerns continuity. Meta’s model card tests recall when conversation details change, alongside persistence through obstacles in personal-agent tasks. The familiar comparison is a project notebook: storing every meeting is easy; finding the revised requirement when it changes the next decision is the useful part. Call that remembering the current requirement. A longer record helps only when the agent can distinguish an obsolete instruction from its replacement.

The model card tests part of that problem. MemoryBench includes conversations of 64,000–100,000 tokens with distractions, conflicting facts and later updates. It is an internal evaluation, graded by another model. That is stronger evidence than a memory-themed demonstration, but it does not measure months of real projects or prove that stored preferences remain appropriately scoped.

There is also an economic boundary. Artificial Analysis measures Wonder at $4.52 per intelligence-index task, versus $1.60 for Spark. Full has no comparable per-token task price. Capability improved; cheap substitution is a separate proposition. Today’s News briefing examines the distribution strategy behind that access split.

The strongest objection is that a benchmark advance may not survive a different agent environment. That objection narrows the conclusion rather than erasing the result. Wonder belongs in comparisons involving sustained work, with its own endpoint and working context preserved. It has not earned an automatic replacement decision.

The decisive follow-up is a matched project test containing midstream requirement changes: does Wonder finish more tasks correctly than Spark without carrying superseded requirements into the final artifact?

A Correction Is Not a Constitution

Compound learning can preserve the wrong lesson with increasing efficiency. Katie Parrott’s account of rebuilding her writing setup, first published October 5 and updated today, identifies a failure that instruction-length caps alone cannot prevent: local editorial feedback became a permanent obligation.

In her first-person report, one essay accumulated 127 drafts. Repeated corrections became checklist requirements, while cross-links pulled unrelated material into the assignment. She then archived the accumulated material and made access opt-in. Combined voice and style guidance fell from 6,109 words to 513. The replacement explicitly treats drafts and reviews as evidence rather than new instructions. These are reported artifact counts and a reported recovery, not a controlled productivity experiment.

The distinction extends the familiar practice of saving lessons after mistakes. Think of a retrospective: “this customer needed a shorter answer” belongs in the incident record; “every answer must be short” changes the operating policy. Automatic promotion skips the judgment between them. A compact instruction file can still contain the wrong generalization, so reducing its size does not resolve the underlying problem.

There is separate evidence for retaining narrower memories. In Memorable’s October report, a learned procedure reduced a repeated Claude Code fixture from 16 turns to 13, with every run passing. It points to relevant files and a verification command, allowing the agent to skip rediscovery. The vendor’s small, repeated task is a limited test, but its mechanism is useful: retrieve a route through a familiar problem without making that route compulsory for every problem.

The practical refinement is to distinguish a remembered observation, a reusable procedure and a binding rule. For example, a failed migration can remain searchable evidence. Its successful recovery sequence can become a task-specific procedure. A prohibition on deleting production data needs a different status and stronger protection. That separation preserves learning while reducing accidental policy changes.

Parrott’s reset changed several things together, so shorter guidance cannot receive all the credit. Nor would excluding old material suit every research task. My inference is that promotion into permanent guidance needs evidence of recurrence across different assignments, beyond repeated frustration within one assignment.

The test is whether the revised setup reduces revision rounds on unrelated tasks while retaining previously demonstrated safeguards. Fewer words alone would not establish success.

The Contrarian Take

Everyone says: A model that follows instructions more faithfully will need fewer surrounding controls.

Here’s why that’s incomplete: Better obedience can increase the cost of a bad instruction. The engineering implication is that the instruction author becomes part of the system under test. Artificial Analysis’s FrontierCode evaluation makes the distinction concrete: acceptable changes depend on maintainer-authored criteria, including constraints beyond functional correctness. Those criteria encode a judgment about what belongs in the project. My inference is that a stronger model can execute an inappropriate criterion more consistently, making the error harder to notice through ordinary pass rates. An independent review must be able to challenge the criterion itself, not merely grade compliance with it.

Under the Radar

A phone can provide a separate input path. Simon Willison’s October 12 web-edit release lets him edit files while driving a remote coding agent, without sending credential contents through the conversation. This extends September 21’s direct-secret-entry pattern to mobile file editing. The repository documents no authentication and expects a trusted network or Tailscale. The separation avoids one disclosure route; it does not stop an agent with filesystem access from reading the resulting file.
A successful exploit can fail the intended test. Meta’s model card reports 92 captured ExploitGym flags, but only 21 used the intended vulnerability; the other 71 used unintended vulnerabilities. The comparable ranking figure is therefore 21. For agent evaluations, this creates two distinct records: success against the intended challenge and unexpected ways the environment can fail. Discarding the latter from the score is appropriate; discarding it from the engineering investigation would lose useful evidence.

Quick Takes

FrontierCode changes the coding shortlist. Artificial Analysis’s October 12 comparison gives GPT-6.1 Sol a 49% rubric score at $0.81 per task, versus Opus 5.5’s 52% at $10.15, using maximum effort, with fallback enabled for Opus. That puts the incremental quality gain under an unusually demanding price test. The evaluation uses 100 tasks, three attempts each and a common agent harness. Scores include model-judged review criteria; they are neither literal human merge rates nor the cost of a successful production patch. (Source)
Interactive design needs its own model comparison. The new AA-3D evaluation asks agents to build working three.js scenes and uses blind human comparisons. This is directly relevant to generated interfaces and browser experiences: ordinary coding correctness cannot determine whether an interaction is useful or coherent. Panel preference is a different acceptance criterion from passing a functional test. It does not establish accessibility, production robustness or a universal design winner. (Source)
Remembered browser actions need a stop condition. Memorable reports 211 and 215 completed return visits out of 240 on two changed-page sets, versus 174 and 187 without memory. Its replay checks each step and hands control back to a model when the check fails. The useful practice is conditional reuse: a previous successful sequence remains valid only while its assumptions hold. These are vendor-run form tasks, not evidence of comparable reliability across arbitrary websites. (Source)

The Thread

Persistent agents create a new maintenance obligation: the instructions themselves need a lifecycle. My inference is that completion capacity can be consumed by obligations that should have expired. An agent may spend its remaining work reconciling old preferences with the current request, even when the model is capable of executing either one.

Consider a research agent whose owner once demanded a comprehensive market map, then requests a short update. If “comprehensive” became permanent policy, stronger recall can trigger unnecessary research on every update. The useful memory is the earlier map plus the circumstances that justified its depth. The stored output and the continuing instruction have different lifetimes.

This extends yesterday’s completion-reserve argument: preserving capacity also requires retiring obsolete work requirements. Expiration risks losing a valuable lesson; permanent retention risks imposing it everywhere. The revealing comparison holds the model and task budget fixed, varies which remembered rules remain active, and measures accepted output alongside the recurrence of old errors.

Predictions

I predict: By November 13, GPT-6.1 Sol will remain within five rubric-score points of Opus 5.5 on Artificial Analysis’s FrontierCode 1.1 Main comparison while costing less than one-quarter as much per evaluated task, under the published configurations. The current matched comparison supports this narrower claim; a changed benchmark version would require separate resolution. (Confidence: medium; Check by: 2026-11-13)

2026-10-13 · 03:53 AM ET