How much compute does an enterprise burn when an autonomous agent logs every micro-action instead of executing its primary task?
Anthropic examined how its Claude models store information in external note files across sessions and retrieve it for later tasks while the model weights remain unchanged. As the findings indicate, retaining raw context without filtering creates administrative deadlocks, degrading task completion rates. Efficient agent memory depends on selectively pruning low-value data rather than accumulating exhaustive activity logs that consume token capacity.
The Cost of Excessive Logging
When autonomous agents log every operational detail, memory stores expand rapidly. The enterprise is left asking how many tokens are wasted when an agent treats every intermediate observation as permanent knowledge.
In a test conducted by Anthropic using the game Pokémon Red, the older Claude Sonnet 3.7 treated every event as potentially important. Over roughly 10,000 game steps, Claude Sonnet 3.7 generated 413 files across 225 folders, creating separate entries for individual battles and discovered items. Because game environments alter continuously, this archive decayed into stale data. Rather than advancing toward objectives, the agent consumed its compute budget reorganizing folder hierarchies, ultimately securing just a single badge.
The older Claude Sonnet 3.7 treated everything as potentially important and, across roughly 10,000 game steps, created 413 files in 225 folders.
By comparison, the newer Claude Fable 5 adopted a targeted operational footprint. As Anthropic detailed, Claude Fable 5 restricted its memory to a few files, ignored information already visible on screen, and recorded high-level plans, strategies, and verified rules, which allowed it to earn four badges.
Why did rigid documentation rules fail to resolve this organizational bottleneck?
Rigid constraints often degrade agent effectiveness instead of improving it. In Anthropic's benchmarking, rigid instructions such as "only record mistakes" wiped out most of the operational advantage seen in stronger models. Older models still required explicit operational parameters, such as mandatory prompts instructing them to inspect existing files before every output.
Correcting Distortions Through Autonomous Review
Exhaustive logging introduces operational risk when an agent relies on faulty internal notes. In Anthropic's gaming run, Claude incorrectly recorded the geographic placement of a mountain landmark, fell into the identical trap three times based on that note, and stalled. To address corrupt notes, Anthropic implemented a post-session maintenance mechanism called Dreaming, based on research from Letta and UC Berkeley. During idle periods, the agent reviews session transcripts, compares notes against chronological records, and resolves erroneous data.
In an Anthropic test with Pokémon Red, Claude fell into the same trap three times, eventually wanted to give up, and missed its goal.
Beyond correcting localized logic errors, Dreaming maintains structured institutional knowledge. Claude parsed roughly 350,000 internal documents from Stainless, a software company Anthropic acquired, assembling a linked knowledge base. In Anthropic's legal evaluation, Claude scored 87 percent accuracy on cross-document queries when using this structured wiki alongside retrieval, compared to 73 percent when relying strictly on search alone.
As Anthropic announced, the company launched Dreaming in May as a research preview within Claude Managed Agents, supporting Claude Opus 4.7 and Sonnet 4.6 by processing up to 100 historical sessions to remove duplicate and obsolete entries.
Anthropic acknowledges that current memory architectures carry material overhead. Note-taking during live tasks distracts the agent from its immediate operational goals, Dreaming requires supplementary compute cycles during idle windows, and knowledge stores still risk compounding hallucinations without continuous pruning.
The catch
An agent’s drive to record every single observation paradoxically cripples its performance, while selective filtering drives success: logging every step buried Claude in hundreds of folders while yielding just one badge. Instead of working, the system wastes scarce context budget on maintaining its own digital clutter. Will the compute budget spent on offline memory consolidation justify the 87% benchmark accuracy when deployed across enterprise repositories?
Our stories: Anthropic Loop Engineering: New Autonomy Levels for Claude (July 17).