Amazon Bedrock AgentCore Memory announces metadata for long-term memory
AI agents can now tag and filter long-term memories using structured metadata, enabling more precise and accurate information retrieval.
View original announcement →Visual Summary
graph TD
A{{AgentCore Memory Metadata}}:::announced
B((AI Agent)):::external
C([Metadata Schema]):::feature
D([Semantic Search]):::feature
E([Metadata Filters]):::feature
F([Auto Extraction]):::feature
G(Amazon Bedrock LLM):::compute
H(Long-Term Memory Store):::storage
B ==>|"ingests events"| A
A -->|"defines keys"| C
A -->|"stores records"| H
G -->|"infers metadata"| A
A -.->|"extraction instructions"| F
F -->|"uses"| G
B -->|"retrieves"| E
E -->|"filters"| H
H -->|"vector search"| D
D -->|"returns results"| B
classDef announced fill:#ff9900,stroke:#ec7211,color:#fff,font-weight:bold
classDef compute fill:#e3f2fd,stroke:#1565c0,color:#1565c0
classDef storage fill:#e8f5e9,stroke:#2e7d32,color:#2e7d32
classDef feature fill:#fff3e0,stroke:#e65100,color:#e65100
classDef external fill:#f5f5f5,stroke:#616161,color:#616161
What's New
Amazon Bedrock AgentCore Memory has added metadata support for long-term memory (LTM) records, allowing AI agents to tag, filter, and retrieve memories using structured attributes in combination with semantic search. Developers can define up to ten indexed metadata keys per memory resource, supporting STRING, NUMBER, and STRING_LIST data types, along with operator-based filtering at retrieval time. This enhancement enables agents to precisely scope memory lookups using structured attributes such as ticket numbers, priority levels, or dates, improving both retrieval accuracy and response quality.
How It Works
- Developers define a metadata schema directly on the memory resource, specifying indexed key definitions that include a key name, data type (STRING, NUMBER, or STRING_LIST), and optionally a set of allowed values.
- Metadata can be attached to memory events in one of two ways: explicitly at ingestion time by the calling application, or automatically inferred by the underlying LLM using extraction instructions embedded in the memory resource configuration.
- During ingestion, the LLM processes incoming conversation events and applies the appropriate metadata values to the resulting LTM records based on those instructions.
- At retrieval time, agents can combine traditional semantic (vector) search with structured metadata filters — using various operator types (e.g., equality, range, list membership) — to narrow results to only the records that match specific structured criteria, eliminating irrelevant context before it reaches the model's prompt.
Why It's Important
- Pure semantic search over long-term memory can surface contextually similar but situationally irrelevant records, introducing noise into an agent's reasoning context and degrading response accuracy.
- By layering structured metadata filters on top of vector similarity search, this feature enables agents to perform highly targeted memory retrieval — for example, fetching only memories related to a specific support ticket, a high-priority customer, or interactions within a particular date range.
- This is especially significant for enterprise agentic workflows where memory stores grow large over time and precision retrieval becomes critical to maintaining coherent, trustworthy agent behavior.
- The LLM-driven automatic metadata extraction also reduces the engineering burden of manually tagging every ingested event.
How It's Different
- Prior to this release, Amazon Bedrock AgentCore Memory's long-term memory retrieval relied primarily on semantic similarity search, which matches records based on embedding-space proximity but lacks the ability to enforce hard structural constraints on results.
- This update introduces a hybrid retrieval model — combining vector search with structured attribute filtering — analogous to filtered vector search patterns seen in dedicated vector databases like Pinecone or OpenSearch, but natively integrated into the AgentCore Memory service.
- The addition of LLM-inferred metadata generation is a differentiator compared to typical vector store metadata implementations, which require the application layer to supply all metadata explicitly; here, the extraction logic is declaratively configured on the resource and executed automatically during ingestion.
When to Prefer It
- This feature is best suited for agentic applications where memory records are naturally associated with discrete, queryable attributes — such as CRM or ITSM agents that need to recall interactions tied to specific ticket IDs, customer tiers, or issue priorities.
- It is particularly valuable when the memory store is expected to grow to thousands or millions of records, where semantic search alone would return too many loosely relevant results.
- Teams should prefer LLM-inferred metadata when conversation content is rich and structured enough for the model to reliably extract attributes, and should use explicit metadata attachment when the calling system already has authoritative structured data (e.g., a ticket number from an upstream API) that should not be left to inference.
- If retrieval precision is not a concern and the memory corpus is small, the added schema definition overhead may not be warranted.
Availability
- This feature is generally available (GA) as of May 6, 2026, and is available in all AWS Regions where Amazon Bedrock AgentCore Memory is currently supported.
- No preview or waitlist period is indicated.
- Key limitations to be aware of include a maximum of ten indexed metadata keys per memory resource, and type support is constrained to STRING, NUMBER, and STRING_LIST — more complex nested or boolean types are not mentioned as supported.
- Developers should consult the Amazon Bedrock AgentCore Memory documentation for the full list of supported regions, operator types available for filtering, and any service quota details.
