State & memory
Reliable agents need more than chat history. Agent RT separates model context, typed workflow state, persistent memory, checkpoints, files, artifacts, and execution history so each can have its own lifecycle and policy.
Model context
Each model request can combine instructions, selected history, workflow state, retrieved knowledge, files, observations, tools, and runtime metadata. Context selection is explicit; durable state remains separate from what the model sees on a particular turn.
Workflow state
Typed state can hold task variables, plan state, counters, approvals, tool results, artifacts, and execution metadata. This avoids reconstructing operational state from natural-language messages.
Memory scopes
Short-term memory is scoped by stable session and thread identities. Persistent memory adds explicit CRUD/search semantics, retention and expiration, and policies for what should be written or retrieved. Semantic, episodic, and procedural memory can share the same scoped boundary while using different ranking or storage implementations.
Checkpoints and resumption
Checkpoints preserve enough execution state to continue after interruption, including messages, durable workflow state, and already-consumed budgets. Resumed work therefore does not silently reset turn, tool, or token limits.
Files and artifacts
The filesystem abstraction provides controlled workspace operations. Artifacts treat reports, code, datasets, images, and other outputs as first-class objects with revisions, lineage, retention, and provenance.