Retrack vs Letta: memory service or agent platform
Letta is a platform for stateful agents that edit their own memory; Retrack adds verified experience memory to the LangGraph agent you already have.
Updated
Letta and Retrack take different architectural approaches. Letta is a platform for stateful agents: the agent runs inside Letta and edits its own memory blocks as it works, with sleep-time agents consolidating memory between sessions. Retrack adds experience memory to the agent you already have: one line in a LangGraph or LangChain agent, with playbooks and warnings learned from verified runs.
At a glance
| Letta | Retrack | |
|---|---|---|
| What it is | An agent platform and runtime with built-in memory | A memory service for agents you build elsewhere |
| How memory changes | The agent edits its own memory blocks with tools; background agents reorganise memory | A pipeline outside the agent verifies runs and distils playbooks and failure signatures |
| Adopting it | Build or move your agent onto Letta | Add retrack.init to your existing agent |
| Learning signal | The agent's own judgement and reflection | Verified outcomes: tests, business events, feedback, AI judge as fallback |
| Untrusted input reaching memory | The agent decides what to write | Memory firewall: untrusted runs never change active memory directly |
| Regression tests, model migration, holdout | Not built in | Built in |
| Open source | Yes (Apache 2.0) | The connector (MIT); the service is hosted |
Choose Letta when
- you are starting a new agent and want memory to be part of the agent's own reasoning,
- you want long-running, stateful agents with self-managed context,
- you are comfortable running your agent on Letta's platform.
Choose Retrack when
- you already have a LangGraph or LangChain agent and do not want to move it,
- you want memory changed only by verified evidence, with a person in the loop for anything untrusted,
- you need tests, migration checks and measured lift from the same memory.