Video summary
Code with Claude 2026: Memory and dreaming for self learning agents
Main summary
Key takeaways
Overview
Anthropic product manager Mahesh (Platform team) presents “memory” as the next core primitive for frontier self-learning agents, and introduces “dreaming” as a new mechanism to continually improve and organize that memory. Both are released as part of the Managed Agents API:
- Memory: public beta
- Dreaming: research preview
Why Memory Matters for Self-Learning Agents
Models and agents can now run for many hours (up to near days). The remaining challenge is continuous self-learning + context management over long horizons.
Memory is positioned as the primitive that enables agents to learn about:
- Tasks
- Success criteria
- Common mistakes
- Effective and ineffective strategies
- Environments
- Up-to-date codebases
- Files and assets
- Operational context
- Other agents in the same environment
- Share learnings
- Detect issues elsewhere
- Incorporate improvements
Product: Claude-Managed “Memory” (Public Beta)
Key promise: “maximize intelligence by default” and support multi-agent systems, with enterprise-grade control.
How Memory Is Modeled / Updated
Memory is treated as a file system that the model can manage:
- A hierarchy of files Claude can read and write
- Uses familiar coding tools such as bash/grep to keep memory organized and updated
Claude Opus 4.7 is cited as being strong at file system-based memory, especially at:
- Deciding what to store
- Choosing how to structure it
- Determining how to organize and split files
Multi-Agent Scaling Requirements
Memory must work when hundreds to thousands of agents run concurrently and share state.
Key requirements include:
- Permission scopes
- Different memory stores can have different access levels (e.g., org-wide read-only runbooks vs SRE read-write working memory)
- Optimistic concurrency
- Agents use a content hash / precondition check to avoid clobbering each other’s writes
Enterprise Control Features (Production Readiness)
To support operational use, memory includes:
- Version history / audit log
- Full audit trail of memory updates
- Attribution metadata: which agent, which session, when
- Supports rollback/traceability and safer control of updates
- Standalone / portable API
- Customers can integrate memory workflows outside managed agents (e.g., custom curation pipelines, PII scanning, cloning memory externally)
Product: “Dreaming” (Research Preview)
Dreaming is introduced as an out-of-band process to improve memory beyond what individual sessions can learn.
What Dreaming Does
Dreaming runs batch and asynchronously, separate from the active agent task “hot path.”
It analyzes:
- Recent agent sessions
- Transcripts
To:
- Find patterns and shared mistakes
- Identify inefficient strategies (e.g., repeated behaviors across agents)
- Produce organized, up-to-date memory content as a memory update (a diff)
Dreaming can be scheduled via:
- Console/API/cron
- Triggered after tasks complete
Why Dreaming Exists (Limitations It Addresses)
Multi-agent memory can become:
- Siloed to each session/task
- Incomplete, missing learnings found by other agents/sessions
- Less efficient at keeping large shared memory holistic and current
Design Rationale (Important Properties)
Dreaming has several guiding properties:
- Out-of-band
- Enables cross-agent pattern detection that a single agent couldn’t infer from its own perspective
- Avoids added latency to the agent task
- Separates objectives
- Keeps “memory quality improvement” distinct from “task completion”
- Scales via compute
- Similar to “test-time compute” / thinking models: spend more effort to produce better-organized memory
- Similar to search systems: build a high-quality index upfront so later retrieval is efficient
Demonstrated Outputs (Demo)
In an SRE scenario with multiple specialized agents:
- Agents have memory stores like:
- Org-wide knowledge (read-only runbooks/SLOs)
- SRE memory (read-write working memory updated continuously)
Immediate benefit: when a similar alert recurs, a new agent consults prior investigation notes to short-circuit redundant work, improving token efficiency, intelligence, and speed.
Then dreaming runs:
- Updates specific memory notes using transcript evidence, such as:
- “agents were triggered 60 seconds after CPU spikes → likely inefficient retry logic”
- Performs deduplication and curation
- Consolidates repeated entries
- Removes stale information
- Adds a verification/backfill note so future agents can trust the updates
Reported Results
Example customer: Harvey (legal benchmark)
- Using dreaming increased task completion rate for a legal scenario by 6x
“Frontier Memory System” Framing
Mahesh describes a layered approach:
- Memory primitive: real-time read/write during work
- Dreaming: verification, organization, enrichment, and backfill using session transcripts
Together, these steps bridge toward larger-scale knowledge bases that remain current and efficient.
Main Speaker / Source
- Speaker: Mahesh — Product Manager, Platform Team, Anthropic