Talk by Lamis Mukta (Member of Technical Staff, Applied AI team, Anthropic) at AI Native DevCon London 2026. Originally seen as a clip on X: x.com/hrswatigupta/status/2078350782412530021.
Agents don't learn — intelligence doesn't compound, task 50 is the same as task 1. The progression from CLAUDE.md → memory tools → skills → agent-managed memory files is an attempt to fix that, culminating in "dreaming": a second-derivative process which periodically prunes and curates memory, run out of band (e.g. overnight), leading to measurably better agent performance over time.
Dreaming runs a review pass over session transcripts and memory stores, surfacing hypotheses for what should change and pointing to the specific evidence behind each one — leaving a human to decide which changes to apply, rather than auto-applying them.
Topic-by-topic comparison against gdata-server/JSONHTL, Envoy, and MCP as a cross-client substrate, plus proposed action items: Application: Dreaming vs. the Notes System, Envoy, and MCP.
YouTube: Learning While You Sleep: Beyond Memory to Dreaming (AI Native DevCon London 2026)
Transcript of the talk (usetranscribe.io)
Tessl podcast write-up: Inside Anthropic — How Claude Tag Is Changing Agentic Work — covers the dreaming concept in the context of Anthropic's managed agents offering.
Cutover blog: AI Native DevCon 2026 impressions — conference summary framing her argument.
Tessl speaker page / talk abstract
No dedicated Anthropic paper or first-party blog post found as of 2026-07-18 — coverage so far is talk + secondhand write-ups only.