𝗠𝗶𝗹𝘃𝘂𝘀 𝗳𝗶𝘅𝗲𝘀 𝗮𝗴𝗲𝗻𝘁 𝗺𝗲𝗺𝗼𝗿𝘆, 𝗲𝘃𝗲𝗻 𝗮𝗳𝘁𝗲𝗿 𝗔𝗻𝘁𝗵𝗿𝗼𝗽𝗶𝗰'𝘀 ...

Anthropic's Managed Agents solved the crash problem. Memory used to live inside the context window, so every crash meant starting from a blank page; now memory https://t.co/fYAEZAPHs1" / X

Post

Conversation

𝗠𝗶𝗹𝘃𝘂𝘀 𝗳𝗶𝘅𝗲𝘀 𝗮𝗴𝗲𝗻𝘁 𝗺𝗲𝗺𝗼𝗿𝘆, 𝗲𝘃𝗲𝗻 𝗮𝗳𝘁𝗲𝗿 𝗔𝗻𝘁𝗵𝗿𝗼𝗽𝗶𝗰'𝘀 𝗠𝗮𝗻𝗮𝗴𝗲𝗱 𝗔𝗴𝗲𝗻𝘁𝘀. Anthropic's Managed Agents solved the crash problem. Memory used to live inside the context window, so every crash meant starting from a blank page; now memory lives in a session log outside the harness, and wake(sessionId) spins up a new harness exactly where the old one stopped. 𝗛𝗼𝘄𝗲𝘃𝗲𝗿, 𝗠𝗮𝗻𝗮𝗴𝗲𝗱 𝗔𝗴𝗲𝗻𝘁𝘀 𝗱𝗼 𝗻𝗼𝘁 𝘀𝗼𝗹𝘃𝗲 𝘁𝗵𝗲 𝗶𝘀𝘀𝘂𝗲 𝗼𝗳 𝗿𝗲𝗰𝗮𝗹𝗹. The session log can only be sliced by position or timestamp, which lets an agent resume from where it stopped but not search past events by meaning. We ran a code-migration task on Anthropic's Managed Agents. Around step 200, the agent hit a dependency conflict and needed to check whether it had handled a similar one before. The log couldn't answer. The agent's only fallback is loading past events back into the context window for the model to scan, which is expensive, misses matches, and fails entirely once the log outgrows the context window. 𝗔𝗻𝘁𝗵𝗿𝗼𝗽𝗶𝗰'𝘀 𝗼𝘄𝗻 𝗱𝗲𝘀𝗶𝗴𝗻 𝗹𝗲𝗮𝘃𝗲𝘀 𝗿𝗼𝗼𝗺 𝗳𝗼𝗿 𝘁𝗵𝗲 𝗳𝗶𝘅: 𝘁𝗵𝗲 𝗵𝗮𝗿𝗻𝗲𝘀𝘀 𝗶𝘀 𝗮𝗹𝗹𝗼𝘄𝗲𝗱 𝘁𝗼 𝘁𝗿𝗮𝗻𝘀𝗳𝗼𝗿𝗺 𝗲𝘃𝗲𝗻𝘁𝘀 𝗯𝗲𝗳𝗼𝗿𝗲 𝘀𝗲𝗻𝗱𝗶𝗻𝗴 𝘁𝗵𝗲𝗺 𝘁𝗼 𝘁𝗵𝗲 𝗺𝗼𝗱𝗲𝗹. 𝗠𝗶𝗹𝘃𝘂𝘀 𝘀𝗹𝗼𝘁𝘀 𝗶𝗻 𝘁𝗵𝗲𝗿𝗲: • Embed the important events as the agent works • Store the vectors in Milvus • The agent searches its past by meaning With Milvus, millisecond semantic search replaces a 200-step scan. The session log stays exactly as Anthropic built it. Milvus doesn't replace the log, but adds a second way to read it, which is by meaning instead of by position.