Breaking

Test Shows Claude Memory Files Can Bloat Agent Behavior, With Sonnet 3.7 Saving Everything in a Pokémon Red Run

November 7, 2026 at 03:15 EST

AI Agents
Foundation Models

A social media post circulating among AI users claims Anthropic has published details on how Claude's memory works and reports a test in which an older model saved far more information than a newer one, with weaker results in the game Pokémon Red.

Claude memory · Pokémon Red test

Saving everything may not help an AI agent

A reported test pitted an older Claude model that logged every observation against a newer model that stored only off-screen information. The exhaustive approach produced a large, cluttered memory and weaker progress in the game.

413

files saved by Sonnet 3.7

225

folders holding those files

badge earned in the game

Memory stored by Sonnet 3.7

Column height is proportional to item count (files vs. folders, same unit)

413

Files

225

Folders

The folder count is about 55% of the file count, a sign of how much material accumulated.

Sonnet 3.7: saves everything

Records every observation it encounters, producing 413 files across 225 folders. Reported result: one badge.

"Fable 5": saves selectively

Reportedly stores only information it could not see on screen. Its final results are not available in the material reviewed.

Takeaway

If confirmed, memory for AI agents should be curated, not comprehensive.

Based on a single test with no independent replication
Full benchmark details and methodology were not included
Effect size and whether it generalizes beyond this game remain unclear
The "Fable 5" model name could not be independently confirmed

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.