Dynamic Insight AI News - Anthropic releases a study on Agent long-term memory, detailing Claude's 'Dreaming' mechanism. As Agents continuously perform tasks, they accumulate vast amounts of conversation logs, operation records, and notes. However, this information may become outdated or contradictory, and may not be easily accessible when needed.

Anthropic divides memory into four layers: raw records, curated notes, interconnected Wiki knowledge base, and a brief index automatically loaded at each startup. Claude can jot down important information during task execution. 'Dreaming' occurs during task gaps, reviewing historical records to correct errors, remove outdated content, and distill new insights from multiple tasks. Upon next startup, Claude only needs to load the concise index and retrieve relevant memories as needed, without re-reading the entire history.

Anthropic conducted experiments using approximately 350,000 internal Stainless company documents totaling 420 million tokens. Researchers had Opus 5 go through 128 rounds of conversation, continuously reading files, building a Wiki, merging duplicates, and verifying facts.

In response to 25 questions about these documents, using a search tool alone scored 53%, while adding a progressively refined Wiki improved performance to 61%. Answering each question also saved an average of around 20,000 tokens. The Wiki’s advantage was especially evident when answers required synthesizing multiple contracts; however, if the answer resided in a single document, standard search was sufficient.

Anthropic also had Claude play Pokémon Red. The older Sonnet 3.7 version recorded 413 memory files but only earned one badge. The newer Fable 5 version retained just 10 files and earned four badges. The updated model is better at distinguishing between valuable strategies worth preserving and game states that can be rechecked anytime.

Dynamic Insight Beating AI News: Meta announces that personal agents Muse and Meta AI are now using a new batch of specialized models. These models are distilled from the flagship Muse Wonder released in October, acquiring two core capabilities: Active Memory and Active Focus.

Active Memory enables the AI to determine which information remains relevant. The model can distinguish between long-term preferences, temporary requests, and expired data, proactively retrieving relevant content when needed in subsequent tasks. According to Meta's internal testing, Muse correctly uses critical context in 90% of tasks—up from 76% in the previous version. In terms of error rate, this translates to a reduction from 24% to 10%, representing a 2.4-fold improvement.

Active Focus allows the AI to persist despite obstacles. When tools fail, task conditions change, or users modify their requirements mid-task, the model attempts alternative approaches, adjusts its execution plan, and seeks user confirmation when necessary. Meta reports that in everyday tasks encountering unexpected obstacles, Muse’s performance improved by 67% compared to the older version. It can also adapt its strategy when conditions shift, determining when to proceed independently and when to involve the user.

After this upgrade, Muse’s score on Meta’s self-developed "Personal Superintelligence Index" rose from 38% to 47%.