Unverified50% confidenceFactExact time
Anthropic's Dreams mechanism runs during a "dreaming" phase between sessions, reads the existing memory store and 1 to 100 past transcripts, and writes to a separate, reorganized new store; the original store is never modified.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
11/10/2026
First Seen
Valid until: 2/8/2027
Sources
Anthropic's Agent Memory Architecture: The Memory Philosophy of Less Is More
youtubeCloud Codes11/9/2026
Related Entities
Related Claims
UnverifiedThe memory system spans multiple turns and even multiple sessions, persisting information that needs to be retained long-term (user preferences, task progress, past conclusions) to external storage and then recalling it on demand to inject into the context, which is a long-term state problem. The memory system and context management have different responsibilities.70% similarUnverifiedMemory leaves traces in LLM systems at multiple levels—including the explicit storage layer, summary layer, context cache layer, and indirect reasoning layer—forming memory residuals.69% similarUnverifiedNextRole's memory design is divided into semantic memory (AGENTS.md), procedural memory (SKILL.md, loaded on demand), and episodic memory (MEMORY.md), and in the hosted version these three types of memory are saved separately per user.67% similarUnverifiedThe paper proposes shared selective persistent memory, a memory architecture specifically designed for agent systems.67% similarUnverifiedMemCurrent advocates separating memory from context engineering and making it an independent storage and retrieval layer, with the model responsible only for reading memory when needed, not for remembering.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1155598API
curl https://kongchang.com/api/v1/knowledge/claims/1155598MCP
get_claim(id=1155598)