How to Add Long-Term Memory to Anthropic's Managed Agents with Milvus

This post introduces two novel patterns for integrating Milvus with Anthropic's Managed Agents to provide long-term memory. The first pattern enables semantic recall within a single agent session by dual-writing indexed events to a Milvus collection, allowing agents to query their own past decisions and actions using vector search. The second pattern establishes a shared agent memory pool across sessions by partitioning a Milvus collection by tenant, enabling parallel agents to learn from collective experience. The post provides code examples for embedding events, performing session-scoped searches, and utilizing Milvus partition keys for tenant isolation. It also discusses production considerations such as indexing strategy, consistency boundaries, and embedding cost control.