Claude Sonnet 3.7 saved 413 files in Pokémon Red – and ended up with one badge
When AI agents need to remember things from one session to the next, the temptation is to archive everything. Anthropic has tested the opposite strategy – and found that selective notes clearly beat large archives. In a Pokémon Red playthrough, the older model Claude Sonnet 3.7 accumulated 413 files across 225 folders and ended up with one badge. The newer Claude Fable 5 stuck to a handful of files and won four. Behind the numbers lies a concrete principle: what is already on screen does not need to be remembered – what matters are the plans, strategies and rules the agent has verified itself.
How agent memory works
In a research post, reported by The Decoder on 16 November 2026, Anthropic describes how Claude is able to carry knowledge between sessions while the model itself remains unchanged. The mechanism is simple: the agent writes notes to files in a folder, which it can read back later. The model weights are not changed – the memory consists entirely of text the agent has written to itself.
That makes memory a question of editorial choices rather than hardware or training. What should be noted down? What should be discarded? And when should the notes be read? Anthropic's finding is that these choices matter greatly in practice – and that the intuition to save as much as possible points in the wrong direction.
The Pokémon test: two models, two memory styles
To illustrate the difference, Anthropic set the models to playing Pokémon Red. The older Claude Sonnet 3.7 treated everything as potentially important. Over roughly 10,000 game steps, it created 413 files spread across 225 folders, often a separate file for each battle and each item found.
The problem, according to Anthropic, is that game state keeps changing – the archive quickly becomes outdated. At times, the model spent no playtime on anything but reorganizing its notes. In the end, it was left with a single badge.
The newer Claude Fable 5 took the opposite approach. It kept its memory limited to a few files and earned four badges. Fable 5 did not write down information that was already visible on screen, but instead recorded plans, strategies and rules it had tested itself.
The difference is not just in quantity, but in type. An archive of mutable game state loses value as the state changes. Plans and verified rules remain useful even as the environment shifts. This is the distinction Anthropic highlights as the core insight: memory is about knowing what to forget, not about how much you can store.
Instructions that can level out model differences
The results of the Pokémon test concern the models' own ability to choose correctly. But Anthropic also found that the instructions governing memory play a large role – and that they work differently depending on the model's strength.
In one benchmark, according to Anthropic via The Decoder, rigid instructions such as "only record mistakes" wiped out most of the advantage of the stronger models. This suggests that strong models need the freedom to judge for themselves what is worth remembering – and that an overly narrow memory rule binds them to a worse strategy.
Older models had the opposite need. They benefited from clear and explicit instructions, such as actually being told to read their notes before every response. A weaker model cannot arrive at good memory practice on its own; it has to be guided there.
For anyone building or configuring agents with persistent memory, this is perhaps the most actionable part of the finding. Memory instructions are not a neutral setting – they interact with the model's capability, and a rule that helps an older model can hold back a newer one.
"Dreaming": when the agent reviews its own notes after the fact
The third element in Anthropic's post is a process they call "Dreaming". The idea builds on research from the company Letta and UC Berkeley and draws on a parallel with the human brain, which processes experiences during sleep.
In practice, Claude goes through transcripts of previous sessions when there is a pause in the work, and checks its notes against them. The mechanism is thus meant to catch errors that have crept into memory – notes that misrepresent earlier sessions, or that have been made outdated by events the agent did not fully register at the time.
What The Decoder reports in detail is the mechanism itself: transcript review during idle periods, with note-checking as the purpose. How thoroughly the effect of the process has been measured is an open question – there is reason not to assume the method's effect has been documented in detail without further evidence.
Why it matters now
The question of whether agents should remember things between sessions has moved from theoretical to practical as agent deployments grow. An agent working on a project over several days – coding, research, customer support – needs some form of persistence. Current solutions range from raw conversation history to structured note systems, and the choices made there affect both cost (how much text is read into the context), speed and the quality of the agent's work.
Anthropic's main point points toward a specific design recommendation: prioritize compressed, curated knowledge over raw logs, and distinguish between information that is available in the environment anyway and information the agent has synthesized itself. The latter – plans, strategies, verified rules – is what actually needs to be stored.
What we don't know
There are significant caveats, and Anthropic, according to The Decoder, is itself clear about the most important one: the comparison between Sonnet 3.7 and Fable 5 covers only one playthrough per model. One playthrough of Pokémon Red is not a statistical basis. The difference in badges may partly reflect the newer model's general skill at the game, not just its memory strategy – the design does not distinguish between these factors.
In addition, every detail in this story rests on The Decoder's account of Anthropic's post; the original source is not immediately available for independent verification. The model names and test figures above thus come from a single secondary source.
These caveats do not change the overall picture, but they put the right precision on it: Anthropic presents a plausible and concrete design principle – selection beats archiving, and memory rules should be tailored to the model's strength – illustrated by a single, at best anecdotal comparison. For practitioners, the point is worth taking away; for research, robust measurements of exactly how large the effect is, and under what conditions it holds, are still missing.
Sources: The Decoder

