CHAPTERS
Why agents need memory: learning across long-horizon tasks
Ravi frames the core problem: agents are tackling longer, more complex tasks, but context management over long horizons remains hard. Memory is positioned as the mechanism that lets agents improve from one task to the next instead of starting from a blank slate each time.
Milestones leading to memory: MCP → Agent SDK → Skills → Managed Agents
A brief timeline connects recent platform releases to the need for persistent learning. Each milestone expanded what agents can do and how reliably they can operate, setting the stage for a dedicated memory layer.
Vision: shared organizational memory and agent swarms
Memory is described as more than per-agent notes—it’s a shared substrate that can coordinate many agents. The end-state is “organizational memory” where multiple agents contribute to and benefit from a continuously evolving shared understanding.
Memory for Claude Managed Agents: what shipped and early customer results
Ravi introduces the launched Memory feature for Claude Managed Agents and emphasizes out-of-the-box intelligence plus enterprise controls. He cites partner/customer outcomes to show practical improvements from persistent memory.
What’s new about this approach: get out of the model’s way
He contrasts older “memory in the harness” patterns (special tools, conventions) with a newer philosophy: leverage what models already do well. The design borrows from the flexibility of Skills and centers memory around files.
Core design choice: memory modeled as a file system
The talk explains why file-system memory works well with modern Claude models: navigation, editing, and organization are strong capabilities. Newer models can decide what to save, how to structure it, and how to represent it for future usefulness.
Multi-agent and multi-session requirements: shared stores and scopes
Ravi shifts from single-agent memory to shared, concurrent usage. The system supports multiple sessions reading/writing the same memory with scoped permissions and hierarchical organization for scalability.
Avoiding conflicts and enabling auditability: concurrency + versioning
To make shared memory safe in production, the system prevents clobbering and supports traceability. Version control and attribution allow teams to review how memory changed and which agent/session caused each change.
Standalone Memory API: CRUD plus enterprise operations
Memory is accessible outside any single agent loop via a dedicated API. This supports teams building in diverse environments and includes operations needed for enterprise governance.
Why “agents writing notes” isn’t enough: local vs global optimization
Ravi describes limitations observed when agents update memory only as they work. Across sessions, repeated mistakes and inefficiencies persist, and memory can become fragmented or duplicative without a global refinement step.
Dreaming: batch refinement that curates and reorganizes memory
Dreaming is introduced as a feedback loop that analyzes sessions and improves memory quality across agents. It runs out-of-band, proposes optimizations, and produces a verified, better-organized snapshot that agents can adopt.
Why out-of-band dreaming matters: cross-agent insights without latency
The architecture choice—separating dreaming from the agent hot path—creates three benefits: cross-session pattern discovery, cleaner objectives, and zero added latency to live task execution.
Demo: SRE agent platform using org-wide memory + task memory
A practical demo shows an on-call/SRE workflow where agents triage alerts and use multiple memory stores. Shared memory enables coordination across incidents—e.g., knowing a fix is already in flight when a similar alert reappears.
Demo: running a dream, inspecting diffs, and improving future triage
Ravi walks through triggering dreaming in the console (also possible via API), examining the run, and reviewing memory diffs. The dream identifies recurring alert patterns (e.g., timing after CPU spikes) and updates guidance so future agents can act sooner and more consistently.
Wrap-up: memory + dreaming as frontier infrastructure for long-running agents
The closing ties the pieces together: shared memory raises the baseline for all agents, and dreaming further improves quality through global refinement. The broader claim is that these systems are foundational for agents that operate over days and continuously expand their understanding of an organization.
