Lamis Mukta - Learning while you sleep: Beyond memory to dreaming - AI Native DevCon June 2026

A talk by an Anthropic engineer covering the evolution of context engineering for AI agents, from simple Claude MD files to autonomous memory systems modeled as file systems. Key production challenges discussed include concurrency (hash-based conflict resolution), versioning, permissioning, and memory staleness. The talk introduces 'dreaming' — an out-of-band, asynchronous batch process where a dedicated agent fleet reviews session transcripts, identifies cross-session failure patterns, and proposes updates to the shared memory store. This separates memory curation from task execution, giving agents broader visibility and dedicated token budget for self-improvement. Practical design principles for scaling these systems to large multi-agent fleets are covered throughout.