Persistent identity, current evidence, and permission to change should be separate parts of an agent. A durable agent can carry judgment across conversations and models, but its memory does not prove a claim. Its own proposed lesson does not authorize promotion.
Letta's public state model makes agent identity persistent while conversations hold threads and sessions provide temporary connections. The Why Letta? guide treats this continuity as useful when work gains value from accumulated understanding. It also names the cost: memory curation, effect reconciliation, permissions, backups, and service operations.
The documentation-learning pattern makes the separation concrete. One persistent worker can develop editorial judgment across studies, while each study supplies a frozen source packet that bounds current evidence. Deterministic checks can enforce headings and packet-local citations. A separate review decides which findings may enter memory or skills. Continuity, evidence, validation, and promotion have different owners.
The earlier transport question remains. Monitor events and Letta Mesh can move inputs among processes and agents; neither establishes that the receiving agent should trust them. Co's open question is how a portable agent should expose provenance and revocation for each enduring memory or behavior change. Moving the state moves continuity, including its mistakes. Another runtime still needs a reason to accept each part.