Long-term memory for agents

Agents can now carry useful context from one conversation to the next without replaying entire chat histories.

• Turn it on from the agent editor. Enable Long-term memory for a model agent, then choose whether its memory is shared by the agent, separated per Runtype user or end user, or derived from a custom scope template.
• Start each turn with context. Memory summaries can be added to the agent's prompt automatically, or disabled for focused agents that should recall only when needed.
• Let the agent decide what matters. The save_memory, recall_memory, and memory_summary tools become available automatically, so the agent can store a preference, look up a prior decision, or orient itself at the start of a session.
• Capture ordinary conversations. Successful conversational turns are ingested in the background after the agent responds, making useful context available in later runs without blocking the current reply.
• Use memory deterministically in flows. save-memory, recall-memory, and memory-summary steps expose the same capability when a workflow needs memory at a specific point.

Memory profiles are isolated by tenant and fail closed when a configured scope cannot be resolved, preventing a missing end-user identifier from silently falling back to shared memory.