Dreaming
An experimental capability introduced by Anthropic for Claude Managed Agents, released as a research preview. It may involve autonomous reasoning, planning, or knowledge integration performed by an agent while operating in a background state.
Timeline (last 90 days)
The content reviewed by Dreaming includes not only conversations between the agent and the user, but also tool-call records and other metadata.
Lamis divides memory work into two complementary paths: in-session memory (the agent reads and writes memory within a single session) and out-of-session "Dreaming" (a separate agent batch-reviews historical session records and modifies the memory store).
The gradual rollout path Lamis outlined is: first use simple means such as Claude MD and skills; when the number of agents grows, run cycles lengthen, and collaboration becomes more complex, introduce production-grade mechanisms such as versioning, hashing, and tiered permissions; finally consider adding out-of-session processes like Dreaming.
Dreaming adopts batch-style reflection: it replays historical sessions offline and periodically to induce patterns before writing back to the memory store; the benefit is decoupling task execution from memory maintenance and gaining a cross-session global view, while the cost is delayed feedback and the need for versioning and audit mechanisms as a backstop.
The Dreaming feature analyzes transcripts and memories across multiple sessions and automatically consolidates and updates the memory store, so that subsequent agents obtain more accurate information.
Currently Dreaming acts only on memory storage, and in the future it may extend to skills, prompts and agent definitions.
"Dreaming" in the agent context usually refers to an agent performing offline reflection, memory consolidation, or strategy optimization while idle.
Anthropic demonstrated a "Dreaming" research preview feature in Claude Console's "flows": periodically reading the agent's memory store and conversation history to produce a reorganized memory store that merges duplicates and replaces outdated entries.
Among the information disclosed so far, it is still unclear how Devin Memory guarantees accuracy, whether Dreaming's pruning mechanism can promptly identify overturned decisions, and how the agent handles memory conflicts.
Devin's Memory belongs to the semantic layer of long-term memory, while Dreaming is closer to metadata management of long-term memory.
9 more timeline events
All Facts (17)
Dreaming reorganizes the index every day, which in essence adapts the index structure to the currently accumulated memory distribution; it is a latency-insensitive offline maintenance task decoupled from online inference.
60%UnverifiedThe content reviewed by Dreaming includes not only conversations between the agent and the user, but also tool-call records and other metadata.
50%UnverifiedThe gradual rollout path Lamis outlined is: first use simple means such as Claude MD and skills; when the number of agents grows, run cycles lengthen, and collaboration becomes more complex, introduce production-grade mechanisms such as versioning, hashing, and tiered permissions; finally consider adding out-of-session processes like Dreaming.
50%UnverifiedDreaming adopts batch-style reflection: it replays historical sessions offline and periodically to induce patterns before writing back to the memory store; the benefit is decoupling task execution from memory maintenance and gaining a cross-session global view, while the cost is delayed feedback and the need for versioning and audit mechanisms as a backstop.
50%UnverifiedLamis divides memory work into two complementary paths: in-session memory (the agent reads and writes memory within a single session) and out-of-session "Dreaming" (a separate agent batch-reviews historical session records and modifies the memory store).
50%UnverifiedThe Dreaming feature analyzes transcripts and memories across multiple sessions and automatically consolidates and updates the memory store, so that subsequent agents obtain more accurate information.
50%Unverified"Dreaming" in the agent context usually refers to an agent performing offline reflection, memory consolidation, or strategy optimization while idle.
50%UnverifiedDevin's Memory belongs to the semantic layer of long-term memory, while Dreaming is closer to metadata management of long-term memory.
50%UnverifiedDevin's Dreaming is a process that runs asynchronously every day to optimize Devin's index of user memories, rather than optimizing the memory content itself.
50%UnverifiedDevin's memory system consists of two mechanisms: Memory and Dreaming.
50%UnverifiedDreaming is a background process that runs daily, responsible for reorganizing memories, removing duplicate content, linking memories to specific sessions, and distilling new lessons learned.
50%UnverifiedThe 'dreaming' process in Devin's memory system refers to the system periodically reorganizing memories, removing duplicates, and associating them with past sessions and outputs.
50%UnverifiedDreaming is currently in the research preview stage, and the official documentation has not elaborated on its specific mechanisms or use cases.
50%UnverifiedCurrently Dreaming acts only on memory storage, and in the future it may extend to skills, prompts and agent definitions.
50%UnverifiedAnthropic demonstrated a "Dreaming" research preview feature in Claude Console's "flows": periodically reading the agent's memory store and conversation history to produce a reorganized memory store that merges duplicates and replaces outdated entries.
50%UnverifiedCognition has launched Memory and Dreaming features for its programming agent Devin.
50%UnverifiedDreaming is an experimental capability introduced by Anthropic, debuting in the form of a research preview.
50%