Vibeleaderboard
← All Intel
Intel / video

Memory and dreaming for self-learning agents

Source
youtube.com
Author
Claude
Date

Context

Anthropic's Claude Managed Agents already let agents write notes to a memory store, which the product manager in this talk says is modeled as a file system the edits itself. The gap described is that each agent sees only its own session, so duplicates, stale entries and patterns across many agents pile up. Dreaming is a separate batch job, in research preview, that reads a memory store alongside up to 100 past session transcripts and writes a new, reorganized store. Anthropic's documentation says the input store is never modified, so the output can be reviewed or discarded.

In the demo, a job over a week of sessions merged five duplicate entries, removed a stale one and noted that several agents' investigations followed the same 60-second pattern, which no single agent had spotted. The speaker also cites a sixfold rise in task completion for Harvey on one legal scenario, attributed to dreaming, and a 90 percent drop in first-pass mistakes at Rakuten, attributed to memory. Those are customer results relayed by Anthropic, not independent measurements.

Key takeaways · AI-distilled
  • Anthropic's platform PM says managed-agent memory is exposed to Claude as a file hierarchy it edits with bash and grep, and claims Claude Opus 4.7 was state of the art at this kind of file-system memory.
  • To support many concurrent agents, the PM describes per-store permission scopes (for example a read-only org runbook store plus a read-write working store) and optimistic concurrency, where an agent checks a content hash before overwriting memory.
  • The PM describes dreaming as a batch, asynchronous job over recent session transcripts, run on a schedule or after tasks finish; it keeps memory quality as a separate objective from task completion and adds no latency to the agent's hot path.
  • The PM reports customer results: Rakuten cut first-pass mistakes in internal knowledge agents by 90% with memory, and Harvey saw a sixfold rise in task completion on one legal scenario with dreaming in early testing.
  • In the SRE demo, a dreaming job noticed agents kept firing 60 seconds after an upstream CPU spike, which suggested inefficient retry logic. It also merged five duplicate entries into one, removed a stale entry, and added verification notes.
Terms in this piece · Glossary
  • AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
  • MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.
  • benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
Read the source www.youtube.com
More from Claude
Recommended reads
Comments

Checking sign-in…

Loading comments…