AI
Why Memory Is the Next Agent Primitive
Memory is the next primitive after MCP, Claude Code, and Skills. Agents can already run for hours on a task. Between sessions, they reset. Memory changes that: an agent retains task success criteria, common mistakes, and environment context across sessions. Agents can also inherit findings from others working in the same environment. Rockutin’s internal knowledge agents cut first-pass mistakes by 90% after deploying Anthropic’s memory system, which also reduced token costs and latency.
Memory as a File System Claude Manages Itself
Anthropic’s managed agents memory treats storage as a file system Claude controls directly. Claude uses bash and grep to manage those files, the same tools that make it effective at agentic coding. Claude Opus 4.7, launched last month, is state-of-the-art at this approach. It decides what content belongs in memory and how to structure the file hierarchy. Delegating those decisions to Claude removes overspecified tool schemas that constrained earlier designs like CLAUDE.md and the original SDK memory tool.
Scaling Memory Across Hundreds of Concurrent Agents
Anthropic and other enterprises already run hundreds or thousands of agents concurrently sharing the same memory state. Two properties emerged from this scale. First, permission scopes: an agent can have read-only access to an org-wide runbook while holding read-write access to a task-specific working memory. Second, optimistic concurrency: before overwriting a file, an agent checks a content hash to confirm no other agent has already modified it. Both properties ship in the current managed agents memory API.
Dreaming Finds Patterns Agents Miss Individually
Dreaming, launching today in research preview in the managed agents API, is a batch asynchronous process that scans recent agent session transcripts across an entire fleet. It looks for patterns no single agent notices from its own context: a failed tool call repeated by many agents, a retry loop triggering exactly 60 seconds after a CPU spike, duplicate memory entries. It produces a diff of updated memory files. Harvey deployed Dreaming on a legal benchmark and saw a 6x increase in task completion.
Why Out-of-Band Processing Improves Memory Quality
Dreaming runs outside active agent sessions, giving it two advantages. It sees across the full fleet rather than one agent’s narrow task context, so it catches systemic patterns. It also separates memory quality as its own objective, keeping task-focused agents from juggling two conflicting goals at once. Because it runs in the background, it adds no latency to active sessions. Murag compared it to test-time compute: spending more tokens upfront produces better outcomes for all downstream agents reading from that memory store.
How Agent Memory Reduces Redundant Work Across Sessions
The demo showed an SRE agent system reacting to P1 alerts by spinning up investigation agents. Each agent carries two memory stores: a read-only org-wide runbook with SLO guidelines and on-call owners, and a read-write SRE memory it updates as it works. When the same alert fires again, the next agent reads the prior agent’s notes and skips redundant investigation. Version history in the Claude console shows every write, the agent that made it, and the content hash used for concurrency checks.
Notable Quotes
So memory is the next primitive. Mahesh Murag · ▶ 2:08
Hey, we already did this investigation. Mahesh Murag · ▶ 19:27
that 60-second pattern every single time. Mahesh Murag · ▶ 22:19
Key Takeaways
- Memory in Claude managed agents models storage as a file system, letting Claude Opus 4.7 organize notes using bash tools.
- Dreaming scans agent session transcripts out of band, finds fleet-wide patterns, and updates memory without adding latency.
- Rockutin cut first-pass mistakes by 90%; Harvey saw a 6x task completion gain after deploying memory and Dreaming.