Claude Sonnet 4.5 and the memory Omni-tool in Letta (www.letta.com)

Letta now integrates Anthropic’s memory tool into its agent platform, letting agents use create/read/update/delete operations on persistent “memory blocks” directly — no custom tooling required. Claude Sonnet 4.5, released alongside this feature, is publicly post-trained to be context-aware and to call the memory tool effectively, so Sonnet agents can dynamically create, delete and reorganize memory blocks as they learn. Letta exposes a filesystem-like API (paths and block_id) that maps those files to Letta’s memory blocks, visualizes them in the Agent Development Environment (ADE), and keeps memory model-agnostic, shareable and accessible via the Letta API for collaborative agent workflows. This matters because it advances agentic memory management beyond simple retrieval/RAG by letting agents edit their own context state to avoid “context pollution” in large context windows. Sonnet 4.5’s post-training reduces the need for heavy in-context prompting, while Letta’s storage abstractions (blocks, filesystem, archives) let any model on the platform leverage the tool. Developers retain manual control (query/edit/share blocks) and can attach memory blocks between agents for transfer learning. The feature is available in Letta Cloud and pairs with Letta’s benchmarking work (Letta Leaderboard, Recovery-Bench), signaling a growing ecosystem for evaluating robust, self-managing agent memory.