Anthropic on agent memory: "dreaming" over 128 sessions lifts a wiki-only score from 13% to 48%, and wiki plus search saves about 20,000 tokens per question
Anthropic breaks agent memory into parts, writing in the moment versus consolidating offline by "dreaming" over transcripts, and reports what each part did on its own evaluations.
// Key points
- Building a wiki from about 350,000 internal files (roughly 420M tokens) with Opus 5 agents, scaling dreaming from 2 to 128 sessions raised the wiki-only score on 25 questions from 13% to 48% of graded criteria
- With search plus the wiki the score rose from 54% to 61% as the wiki grew, versus 53% with search alone, and saved about 20,000 tokens per question; on 24 merger questions the wiki scored 87% versus 73% for search
- In Pokémon, Sonnet 3.7 wrote 413 memory files across 225 folders and earned one badge, while Fable 5 kept a small folder of strategy and tested rules and got much further
- Telling Opus 5 and Opus 4.7 to "record only failures" cost them much of their memory gain; prompting Sonnet 4.6 to read memory and test its theories nearly doubled its score
Builder's takeThe useful finding is that what to remember matters more than how much: Sonnet 3.7's 413 files made it slower, not smarter. For the AI cloud drive I'd stop storing anything the user can re-read from the file itself, and run a nightly consolidation pass over the day's logs instead of writing memory mid-task.
// Background · from #Anthropic
Full timeline →- Nov 7 AWS runs Claude Code as a remote agent on Bedrock AgentCore: one microVM per task, up to three attempts, and a draft pull request unless verification passes
- Nov 6 Claude Console adds an organization IP allowlist in beta, up to 1,000 addresses or CIDR ranges, and the Admin API can sign an Enterprise member out everywhere
- Nov 4 Claude's Messages API adds message threads in beta: later turns send only the new messages, and a thread expires 5 minutes after its last request