Claude series models have started independently structuring their own experience for more efficient task execution
Artificial intelligence Claude has gained new capabilities for self-improving its memory during periods of inactivity. The models are learning to self-correct errors, structure notes, and filter out irrelevant information without external prompts. This was reported by Anthropic, according to the portal PromPolitInform.
Developers have implemented an algorithm dubbed “dreaming,” drawing an analogy to memory consolidation processes in the human brain. During pauses between user interactions, the system analyzes past logs and transcripts to verify the relevance of stored rules and refine accumulated data.
Knowledge update mechanism
Instead of storing every step, as the Sonnet 3.7 model did, more advanced versions, such as Fable 5, have learned to evaluate the significance of events. This prevents memory saturation with outdated or redundant information that constantly changes in dynamic environments, such as during gameplay or complex task execution.
Research based on CL-Bench 1.0 testing demonstrated that Opus series models are capable of forming cohesive strategic documents. They not only record errors but also create rules to avoid them in the future, closing the learning loop without direct operator intervention.
Autonomous data filtering
Developers discovered that providing excessive instructions on exactly what to store in memory can reduce the efficiency of intelligent systems. The best results are demonstrated by models that are given the freedom to independently determine the priority of information.
Dreaming allows Claude to identify inaccuracies that arise during sessions and update its knowledge base. This improves system reliability as it verifies its previous assertions regarding current tasks before utilizing them in new work sessions.
Illustrative photo: Alan Warburton / CC BY 4.0
Read also: Anthropic prohibits creating game clones in Claude: what is the reason
