Summary
Letta (formerly MemGPT) is a stateful agent execution framework with an official Python client SDK, maintained at https://github.com/letta-ai/letta with docs at https://docs.letta.com. Sending a message to a running Letta agent triggers multi-step reasoning, tool calls, and memory operations — an agent-run execution surface comparable in scope to other agent frameworks this repo already instruments (crewai, autogen, agno, strands, agentscope).
This repo does not instrument Letta anywhere: no py/src/braintrust/integrations/letta* directory, no wrapper, no nox session, no cassette, no README entry, and no mention in py/pyproject.toml's [tool.braintrust.matrix].
Braintrust docs status: not_found. https://www.braintrust.dev/docs/guides/tracing does not mention Letta or MemGPT among its supported frameworks list.
What needs to be instrumented
The primary execution entrypoint is sending a message to an agent:
client.agents.messages.create(agent_id, input=..., messages=[...]) — synchronous, POST /v1/agents/{agent_id}/messages- streaming variant (
streaming=True, SSE) for token/step streaming
The response is a LettaResponse containing a messages array of typed steps (ReasoningMessage, ToolCallMessage, ToolReturnMessage, AssistantMessage, etc.), usage (LettaUsageStatistics), and stop_reason — a rich multi-step agent trace analogous to what other agent-framework integrations in this repo already capture as nested spans.
Upstream sources
Local repo files inspected
py/src/braintrust/integrations/ — no letta directory- Repo-wide grep for
letta (case-insensitive) under py/src/braintrust/ — zero matches
README.md integrations table — no Letta/MemGPT entrypy/pyproject.toml [tool.braintrust.matrix] — no letta key
Summary
Letta (formerly MemGPT) is a stateful agent execution framework with an official Python client SDK, maintained at https://github.com/letta-ai/letta with docs at https://docs.letta.com. Sending a message to a running Letta agent triggers multi-step reasoning, tool calls, and memory operations — an agent-run execution surface comparable in scope to other agent frameworks this repo already instruments (
crewai,autogen,agno,strands,agentscope).This repo does not instrument Letta anywhere: no
py/src/braintrust/integrations/letta*directory, no wrapper, no nox session, no cassette, no README entry, and no mention inpy/pyproject.toml's[tool.braintrust.matrix].Braintrust docs status:
not_found. https://www.braintrust.dev/docs/guides/tracing does not mention Letta or MemGPT among its supported frameworks list.What needs to be instrumented
The primary execution entrypoint is sending a message to an agent:
client.agents.messages.create(agent_id, input=..., messages=[...])— synchronous,POST /v1/agents/{agent_id}/messagesstreaming=True, SSE) for token/step streamingThe response is a
LettaResponsecontaining amessagesarray of typed steps (ReasoningMessage,ToolCallMessage,ToolReturnMessage,AssistantMessage, etc.),usage(LettaUsageStatistics), andstop_reason— a rich multi-step agent trace analogous to what other agent-framework integrations in this repo already capture as nested spans.Upstream sources
Local repo files inspected
py/src/braintrust/integrations/— nolettadirectoryletta(case-insensitive) underpy/src/braintrust/— zero matchesREADME.mdintegrations table — no Letta/MemGPT entrypy/pyproject.toml[tool.braintrust.matrix]— nolettakey