Memory and context
Agent memory supplies history and knowledge to later model calls. A model does not remember earlier requests unless that information is included again as context.
Session history and domain knowledge
Session memory holds the history of related interactions, identified by a session ID. The runtime stores and loads it using durable entities, so a conversation can continue after a restart. Agents using the same session can share history, with filters controlling which entries each agent sees.
Domain knowledge, such as customer preferences or product data, belongs in your application’s entities, views, and external data sources. Tools retrieve relevant information and add it to the current request. Retrieval-augmented generation uses this approach to supply knowledge outside the model’s training data.
Long-term memory carries knowledge across sessions, users, and agents. It is distinct from the history of one conversation and from the domain data your application owns.
Keeping context useful
More history increases token cost and can make important information harder for a model to identify. Limit or filter the entries supplied to each call. Compaction replaces older history with a shorter representation, trading detail for a smaller context. Sequence-aware compaction preserves messages added while the summary was being prepared.
When splitting work across agents, choose context boundaries deliberately:
-
Focus: include what the current task needs.
-
Relevance: remove obsolete intermediate work.
-
Isolation: keep unrelated concerns separate.
-
Independence: avoid sharing earlier reasoning when you want diverse answers.
An Autonomous Agent has private execution history. Delegation, handoff, teams, and moderation share different subsets of it; they do not automatically share every model interaction.
Next steps
-
Session memory configures filtering, compaction, interception, and replication.
-
Context across agent boundaries describes each capability.
-
Executing RAG queries retrieves external knowledge.