Oracle Select AI Agent adds configurable memory and team APIs
Oracle released Select AI Agent memory and interoperability enhancements on September 15, 2026, according to Oracle documentation updated October 13, 2026. The features deploy on Autonomous AI Database on Dedicated Exadata Infrastructure and introduce configurable short-term and long-term memory lengths alongside new APIs for agent tooling and team management.
Memory Configuration and Agent Tool Control
The `DBMS_CLOUD_AI_AGENT` package now supports configurable memory depth, with a documented default for `short_term_memory_length` set to 30 conversation turns Oracle documentation. Developers can adjust both short-term and long-term memory parameters to fit application requirements. The release also includes APIs to describe and run tools, allowing agents to execute external functions and integrate with downstream systems.
Agent-Team Discovery and Run-State Management
The enhancement adds APIs to discover agent teams, inspect team capabilities, check team run state, and set the execution state for team runs waiting for human input Oracle documentation. The `DBMS_CLOUD_AI_AGENT.RUN_TEAM` procedure can start a new agent-team run or resume a paused run, enabling orchestration patterns where human operators can pause, review, and resume agent execution mid-workflow.
Team Sharing and Multi-User Deployments
Oracle documentation describes team sharing capabilities that allow a team owner to define and maintain agents, tasks, and tools in one database schema while another database user or role runs the team using an owner-qualified team name Oracle documentation. This architecture separates agent definition (privileged) from execution (delegated), reducing the surface area for permission misconfiguration in multi-tenant or departmental deployments.
Product Positioning
Select AI Agent remains positioned for enterprise deployments on Oracle's dedicated Exadata infrastructure. The memory and discovery enhancements address practical operational challenges: agents must retain context across conversations without unbounded storage, and teams running multiple agents require visibility into each agent's capabilities and execution status. These are foundational capabilities for production workloads rather than experimental features.
The release does not include changes to model selection, fine-tuning, or cost controls. Oracle has not announced pricing adjustments or availability timelines for Standard or Exadata Cloud@Customer deployments.