What’s New in Oracle AI Agent Memory Release 26.6
This article highlights the key changes in Oracle AI Agent Memory Release 26.6.
Features and Improvements
-
Hybrid search for Oracle DB-backed stores
Hybrid search is now available for Oracle Database-backed stores, combining semantic and keyword retrieval in the same search flow.
For more information, see Use Oracle Agent Memory Hybrid Search.
-
Background memory extraction
Background memory extraction is now available as a new processing mode. This enables lower-latency ingestion while extraction continues asynchronously.
For more information, see Agent Memory API Reference.
-
Memory extraction custom instructions
Automatic memory extraction can now be shaped with custom instructions. This enables more targeted memory capture for specialized use cases.
For more information, see Customize Automatic Memory Extraction.
-
More powerful metadata filtering and search
Metadata filtering has been expanded to support more precise search conditions, including richer matching for structured metadata.
For more information, see Use Metadata and Metadata Filtering.
-
Records can now be updated after creation
Update support has been added for existing threads, memories, and messages. This enables stored records to be corrected or refined without recreating them.
For more information, see Agent Memory API Reference and Thread API Reference.
-
Time-to-live for messages and memories
Messages and memories can now expire automatically based on schema-level retention policies or per-record time-to-live settings. This enables applications to manage data retention, including automatic purging of expired records.
For more information, see Use Time-to-Live for Messages and Memories.
-
Broader metadata support across messages, memories, and profiles
Messages, memories, and profiles now support broader metadata storage. This enables richer labeling and filtering across more record types.
For more information, see Use Metadata and Metadata Filtering and Agent Memory API Reference.
-
Automatic extraction can now inherit message metadata
Extracted memories can now inherit message metadata automatically so that metadata-based search and filtering stays aligned with the original message context.
For more information, see Use Metadata and Metadata Filtering.
-
Context card minimum results by type
Minimum result counts can now be set by record type in context cards, which enables more balanced context assembly across preferences, facts, guidelines, and other memory types.
For more information, see Customize Context Card Content.
-
Managed schema upgrades on startup
Supported managed schema upgrades can now be applied automatically during startup when needed.
For more information, see Stores and Schema API Reference.
Deprecations and Compatibility Changes
-
Extraction configuration is moving to
MemoryExtractionConfigInline extraction parameters are now deprecated in favor of
MemoryExtractionConfig.For more information, see Agent Memory API Reference.
-
Python 3.14 is no longer supported
Supported Python versions are now limited to 3.10 through 3.13.
For more information, see Frequently Asked Questions (FAQs) and Troubleshooting.