Inside Anthropic's Agent Memory Architecture: From CLAUDE.md to AI Dreaming
新智元 · wechat · 2026-08-01
An Anthropic engineer recently shared deep insights at DevCon into the company's internal practices for building AI agents that get smarter over time.
Four Generations of Memory Systems
- CLAUDE.md: Markdown files injected at the start of sessions. Highly effective but prone to bloat.
- Memory Tool: Grants agents autonomy to decide when to read, write, and update memory.
- Skills (Progressive Disclosure): Agents load deep details on-demand based on file summaries, much like pulling a book from a shelf.
- File System: Memory is structured directly as a file system with guardrails like rollback, collision prevention, and permission isolation to manage concurrent multi-agent modifications.
The "Dreaming" Mechanism
To prevent agents from being distracted during tasks and lacking global context, Anthropic introduced a dedicated review mechanism. After tasks are completed, a specialized agent reviews historical logs and the memory base to identify recurring errors and propose updates for human approval. This significantly reduces future error rates and inference costs.
Workflow Graph Engineering
Traditional linear workflows often suffer from unnecessary waiting. The guide recommends using "fake edge tests" to remove redundant data dependencies, transforming processes into a parallel "diamond" shape with quality checks before aggregation, drastically minimizing idle time.
More from coding & agent
- Human-AI agent team handles 39 emails, helps student robotics team win 2nd place — toolstelegraph · 2026-08-01
- App Store Connect CLI 3.4.0 released: Optimized session recovery and password management — rudrank · 2026-08-01
- AI Agents Wrote Two Papers in 6 Days for $3K, Both Rejected for Lack of Judgment — rohanpaul_ai · 2026-08-01
- How AI Blurs Job Boundaries: Non-Technical Folks Coding Becomes a Trend — Ok-Airline-8523 · 2026-08-01
- Reddit Thread: Why Non-Coding AI Agent Use Cases Are Mostly Ineffective — chkbd1102 · 2026-08-01
- AI Agents and Closure: Trusting the Workflow — rudrank · 2026-08-01