Agent Memory & State
How should I handle non-replayable agent memory in Catalyst workflows?
Offload non-replayable agent memory outside Catalyst’s durable execution runtime to avoid breaking workflow replay consistency. Catalyst’s durable execution only persists state that can be fully replayed during workflow recovery, so temporary LLM caches, ephemeral draft inputs, or transient session data do not belong in checkpointed workflow state. Do not tie critical workflow progression markers to this unpersisted, non-checkpointed memory that cannot be restored after workflow restart.
Was this article helpful?
Your feedback helps improve Diagrid's FAQ experience.
Keep reading
More Diagrid FAQ articles
- Agent Memory & State
How do workflow state, conversation memory, and retrieval corpora differ for AI agents?
Explain how workflow state, conversation memory, and retrieval corpora differ for AI agents, including their unique roles in durable execution systems.
- Agent Memory & State
Which agent memory and state should be included in durable execution?
Outline which types of agent memory and state are appropriate for inclusion in durable execution layers for production AI agent workloads.
- Agent Memory & State
What agent state survives a workflow crash or restart?
Explain which agent state is preserved following a workflow crash or unexpected restart, and how Catalyst manages state persistence.
