AI News/Wire summary

Anthropic says selective notes beat full logs for Claude agent memory

Key points

  • Pokémon Red test: older Sonnet 3.7 created 413 files and earned one badge, while newer Fable 5 used few files and earned four.
  • Dreaming reviews past transcripts during idle time and corrected a wrong landmark note that had sent Claude down a failing path.
  • Dreaming launched in May as a research preview for Claude Managed Agents, reading up to 100 past sessions without altering the original store.
  • In a legal test, wiki plus search scored 87 percent on multi-document questions, versus 73 percent for search alone, according to Anthropic.

Why it matters for marketers: Teams building agent workflows should favor curated, corrected memory over full logs, though the evidence comes only from Anthropic's own tests.

Summary by the Z Media news desk, from the outlets' reporting. Read the full stories at the source.

Read the original

The Decoder: Anthropic says the key to agent memory is knowing what to forget, not how much to store

More coverage

Related on Z Media