Anthropic: let the model pick what to remember, not fixed rules

Anthropic's engineering team published a post on 6 November 2026 describing how it tests memory for Claude agents, and its main finding is practical: on its own benchmarks, stronger models did better when left to decide what to remember than when given fixed rules. The post is vendor-published and the evaluations are Anthropic's own, so treat the figures as the company's results, not independent measurements.

Let the model judge what to save

On Continual Learning Bench (CL-Bench) 1.0, a database question-answering task run on Claude Managed Agents, each question started a fresh session with only a persistent memory folder carried over. Anthropic reports that Claude Opus 5 was the first of the models it tested to close the learning loop: its notes held a rule, how it knew the rule was true, earlier answers and warnings about traps. Claude Opus 5.5 and Claude Sonnet 5.5 did the same.

When Anthropic told models to record only failures, both Opus 5 and Opus 4.7 lost much of the gain that memory gave them. Older models needed more help: prompting Claude Sonnet 4.6 to read its memory files and test its theories nearly doubled its score, according to the post. Sonnet 5 read memory before answering 9 times out of 10 unprompted; Sonnet 4.6 did so less than half the time.

Smaller memory beat more memory

In a Pokémon Red run about 10,000 steps in, Claude Sonnet 3.7 had written 413 memory files across 225 folders and earned one badge. Claude Fable 5 had written nine files in two folders plus one global.md file and earned four badges. Anthropic says the older model stored facts the screen already shows each turn, which went stale and cost time to search.

'Dreaming': a background pass that fixes notes

The post uses 'dreaming' for a background job in which Claude reviews past session transcripts to correct, reorganise or promote notes. In one Pokémon case a wrong note about where a building stood sent Claude over a ledge three times; after a dreaming pass checked the note against the transcript, Claude went back into the cave and reached the next city.

For document work, Anthropic had Opus 5 agents build a wiki over about 350,000 internal files (roughly 420 million tokens) from Stainless, a company it acquired in May. Across 128 sessions, scores on 25 questions rose from 13% to 48% of graded criteria using the wiki alone. Search plus the wiki reached 61%, against 53% for search alone, and saved about 20,000 tokens per question.

The wiki did not always help. On a merger question set, Claude scored 87% with the wiki against 73% with search on the questions needing several documents. On a task where one document held each answer, there was no clear gain, and on another the wiki grew larger than the source files.

Instructions for the wiki mattered

In a contract task, several agents wrote to one shared table and overwrote each other: only 41 of 57 contract facts (72%) survived. After Claude rewrote the instructions so each contract got its own page, written by one agent and quoting terms word for word, 57 of 57 survived.

What the post leaves out, and the pushback

The post gives no cost for dreaming compute, and it says itself that continuously checking every memory may not be practical. Several figures come from single runs. In the Hacker News thread, where the post drew 38 points, a top comment objected that calling saved markdown files that are re-injected into context 'dreaming' overstates what happens. Another commenter said they switch off automatic memory in every tool and delete most of what it writes because it is specific to one task.