How should you design an agent's memory? Anthropic compares it to how humans consolidate memories during sleep, and proposes "dreaming" 🧠 The components of agent memory ❓ What layers does memory have? 💡 From the bottom: transcripts of everything, memory notes written while working, structured representations such as wikis, and a small injected context loaded into every session. A wrong note in the injected layer is carried into every session, so it should stay as small as an index. ❓ What is dreaming? 💡 An offline consolidation process between sessions that reads the lower layers and updates the upper ones. It checks notes against transcripts to fix errors, merges duplicates, and promotes what matters. In Pokémon Red's Mt. Moon, the agent reached Cerulean City after a wrong note was corrected. ❓ Does it actually work? 💡 ・On CL-Bench, Opus 5 rose from 0.564 to 0.828 ・On about 350,000 Stainless files, wiki-only accuracy climbed from 13% after 2 dreaming sessions to 48% after 128 ・Combining wiki and search saved about 20,000 tokens per question Most results come from single runs, so treat effect sizes as indicative. ❓ What should the agent write down? 💡 What it can't recover from the environment: strategies and tested rules, not what the screen or tools show every turn. Fixed rules like "record only failures" shrank the gain for stronger models. #AIAgents# #AgentMemory#