Anthropic has introduced a concept called "dreaming" to address how AI agents manage memory over time, according to a presentation by Lamis Mukta, a member of the company's technical staff, at AI DevCon in London this summer. Dreaming is an asynchronous batch process that runs separately from active tasks to consolidate memories, remove outdated information, and identify recurring failure patterns like incorrect units or broken tool configurations. The approach aims to solve a bottleneck where agents must split their focus between executing tasks and curating their own memory systems.
Anthropic's memory journey began with CLAUDE.md, a file Claude reads at the start of every conversation to establish persistent context through Bash commands, code style, and workflow rules. The company then explored memory tools that let agents autonomously decide when to read, write, and update memories, using progressive disclosure so agents scan light metadata first before retrieving full content and source files. Mukta compared this to "a bookshelf in my room" where someone can "scan and look at my list of books and see if any of the titles might be relevant to the conversation, and then pick that off the shelf and read it when I need to." But human-agent collaboration remained a limiting factor, and memories could grow stale or even be written incorrectly or maliciously injected. Dreaming addresses these issues by spotting patterns where agents consistently fail and proposing memory updates for human review.
Independent AI agent systems researcher Jayakumar Ramalingam warns that "a bad answer normally dies with the session; a bad memory can influence thousands of future sessions," arguing that dreaming creates a dangerous path from repeated mistakes to persistent policy. Enterprise AI architect Vladimir Beskorovainyi tells The New Stack that "the failure that actually costs you is an agent reaching for the wrong tool for a reason that looked perfectly defensible at the time," and cautions that if updates come from recent batches, the memory store drifts toward whatever the agent fleet happened to do lately. Penetration testing company founder Viktor Bulanek adds that dreaming "points an automated writer at session transcripts, and transcripts are full of content the agent did not author," while noting that "a confident note about a system that was refactored last month is worse than no memory at all" because versioning shows what changed, not what rotted.
The technical choice to use filesystem-based memory rather than vector databases drew praise from experts who value legibility and auditability. Beskorovainyi notes the industry spent two years insisting memory meant embeddings, and Anthropic solved it with a filesystem and grep, making it a memory store developers can open and read rather than debug through opaque embeddings. Noah Labs CTO Berk Yilmaz calls the approach's lack of flair "endearing," observing that once a memory decision is made on which past behavior should become future behavior, memory itself stops being inert—a hallucination that survives a thousand sessions becomes infrastructure, and if an attack writes to memory, it becomes persistent. Former Meta data science leader Kerstin Frailey raises cost concerns, noting that dreaming requires paying an AI to do work once, then paying AIs to regularly review, revise, and restructure it, unlike deterministic predecessors like garbage collection. The report notes that OpenAI, Letta, and other platforms are also adopting similar automatic out-of-band memory consolidation approaches, suggesting the industry is converging on the realization that in AI modeling terms, memory is actually maintenance. For organizations deploying autonomous agents at scale, the infrastructure choices around memory versioning, provenance tracking, and rollback capabilities may prove more consequential than the intelligence of the models themselves.