Skip to content
Sıfırıncı Dakika
Latest

Anthropic explores 'dreaming' for Claude: it corrects its memory when idle

AI1 min read

In brief

Anthropic published research on 'dreaming,' a process that lets Claude review past experiences during idle time, correct faulty memories and organize what it has learned. In Pokémon Red experiments, a smaller memory system performed better, and an incorrect location note was fixed through offline review.

  • Reviewing past sessions during idle time
  • Correcting faulty memory notes against the original session transcript
  • Maintaining organized memory with fewer files
  • Building a linked, wiki-like knowledge base from document collections

What should memory hold?

Anthropic says Claude saves useful information to memory files while completing tasks, but recording every interaction creates clutter while skipping important details leaves future sessions without context. The company tested this balance in Pokémon Red.

After roughly 10,000 gameplay steps, two models were compared. Sonnet 3.7 recorded almost everything: 413 files across 225 folders, totaling nearly 978,000 characters. Yet it earned just one badge. Claude Fable 5 kept only 10 files across three locations, using about 43,400 characters, and earned four badges.

How 'dreaming' works

Anthropic's 'dreaming' process draws inspiration from how human brains consolidate memories during sleep. When Claude is idle, it reviews previous interactions, identifies incorrect information, refines its notes and builds structured knowledge that future sessions can access.

The company also showed how faulty memory can be corrected. Claude had incorrectly recorded the location of a Pokémon Center near Mt. Moon, causing it to repeatedly cross a one-way ledge and return to an earlier route. In an offline review, Claude compared its notes against the original session transcript, found the error and updated the file. It then reached Cerulean City, while the version with incorrect memory stayed stuck at the same obstacle.

From documents to a knowledge base

Anthropic also explored whether large document collections could become interconnected knowledge bases. In an experiment with roughly 350,000 internal documents (about 420 million tokens) from software company Stainless, Claude developed a wiki across 128 sessions. Its performance on 25 evaluation questions rose from 13 percent to 48 percent.

Why it matters

Forgetting or misremembering is a core problem for AI agents on long-running tasks. Anthropic's research offers a concrete example of how agents might correct their own mistakes and organize knowledge.

Sources

Related stories