Oracle AI Agent Memory 26.6 Adds Hybrid Search and Full CRUD

Enterprise AI's real challenge isn't conversation—it's memory

Oracle has released AI Agent Memory 26.6, a database-native memory system designed to solve what the company identifies as the core failure mode of enterprise AI agents: not poor conversational ability, but flawed data management. The release introduces hybrid search, full CRUD operations, and lifecycle controls for AI systems operating in regulated, high-stakes environments.

According to Oracle, enterprise agents fail when they retrieve the wrong information, forget critical context, surface data meant for different users, or retain information past its deletion date. Version 26.6 addresses these issues through architectural choices rather than model improvements alone.

Hybrid search combines semantic and exact matching

The new release combines vector search for semantic understanding with keyword search for precise identifiers and exact phrases. The system can recognize that "FY26 close calendar," "FIN_CLOSE_2026_v7," and "fifth business day" may reference the same operational reality—and ensure that reality is scoped to the correct user, agent, business unit, and conversation thread.

Agent Memory 26.6 treats different information types as distinct record classes: customer preferences, financial facts, and safety rules are stored and retrieved as categorically different entities rather than undifferentiated text chunks. Retrieval can be filtered by users, agents, threads, record types, and metadata.

In-database architecture reduces latency

Because Agent Memory runs on Oracle AI Database, embeddings, vector indexes, keyword indexes, metadata, and transactional state coexist in a single system. The OracleDBEmbedder generates embeddings inside the database, eliminating external API calls. HNSW vector indexing and hybrid search execute without moving data between specialized systems.

The release supports vector-only retrieval for speed, keyword retrieval for exact matches, and hybrid search when both are needed. Teams can configure index synchronization and run memory extraction in the background when write operations must return immediately.

Full lifecycle management, including cascading deletes

Version 26.6 adds complete CRUD capabilities for threads, messages, memories, user profiles, and agent profiles. Deleting a thread triggers cascading cleanup of associated messages, memories, and retrieval state. Deleting a user or agent cascades through relevant threads and scoped records.

The system includes time-to-live controls and schema-level retention configuration. Applications can set expiration relative to storage time or event time, addressing privacy requirements and operational governance.

Additional enterprise features include chunked indexing for large content, context cards that combine summaries with recent messages and durable memory, asynchronous APIs for high-concurrency workloads, and metadata filtering for policy-aware retrieval.

Why it matters

As enterprises move AI agents from demos to production, memory management becomes a compliance and accuracy problem, not just a performance optimization. Systems that can't reliably scope information to the right user, enforce retention policies, or execute true deletes create liability in regulated industries. Oracle's approach embeds these controls at the database level rather than layering them on afterward—a architectural choice that reflects how memory requirements differ between consumer chatbots and enterprise operational systems.

These details were first reported by Oracle on the Oracle Database blog.