Source: Rearchitecting Letta's Agent Loop (Letta, 2025)

Letta's letta_v1_agent drops MemGPT heartbeats, the send_message tool and tool-argument inner thoughts in favour of native reasoning and native multi-step tool use.

Contents

Summary

Explains why Letta replaced the MemGPT-style agent loop. MemGPT was designed for models that could not reliably chain tools or reason natively, so it simulated both with tool arguments. Frontier models are now post-trained for multi-step tool use and self-termination, so Letta’s new default loop lets the model drive.

Key claims

  • Removed: the request_heartbeat argument for chaining tool calls; the send_message tool (the model’s plain assistant message is now the reply); the thinking argument for inner thoughts (replaced by native reasoning).
  • Rationale: keeping agents “in-distribution” for models such as GPT-5 and Claude Sonnet 4.5 significantly improves performance.
  • Native reasoning supported (OpenAI via the Responses API; encrypted reasoning handled across providers); tool calling is no longer required to connect a model.
  • Trade-offs: non-reasoning models produce no reasoning and prompted reasoning is unsupported; encrypted reasoning cannot be edited or moved between models; tool rules cannot constrain assistant messages; sleep-time style independent runs need custom prompting.
  • The original MemGPT architecture remains supported; V1 is recommended for current reasoning models.

Feeds

Letta

Suggest changes →