Amazon Bedrock AgentCore Memory now supports strictly consistent metadata for long-term memory
Guarantee exact metadata values on agent memory records — enabling hard multi-tenant isolation and compliance boundaries without LLM inference.
View original announcement →Visual Summary
graph TD
A{{"Strictly Consistent Metadata"}}:::announced
B(Amazon Bedrock AgentCore Memory):::compute
C([Event Grouping Isolation]):::feature
D([No Cross-Value Merging]):::feature
E([Multi-Tenant Isolation]):::feature
F((Application)):::external
G(Short-Term Memory):::storage
H(Long-Term Memory Records):::storage
I([LLM-Inferred Keys]):::feature
F ==>|"attaches metadata"| A
A ==>|"pass-through guarantee"| B
B -->|"ingests events"| G
G -->|"extracts to"| H
A -->|"controls"| C
A -->|"enforces"| D
A -->|"enables"| E
B -.->|"coexists with"| I
I -.->|"enriches"| H
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 now supports strictly consistent metadata, allowing applications to attach metadata values directly to short-term memory events that pass through extraction and consolidation unchanged — with no LLM inference involved. When a metadata key is configured with the STRICTLY_CONSISTENT extraction type, the value supplied by the application is guaranteed to appear on the resulting long-term memory record exactly as provided. This feature also controls how memory events are grouped and consolidated, ensuring records with different metadata values are never merged even if their content is semantically similar.
How It Works
- Extraction type configuration: Metadata keys are declared as indexed keys on the memory resource, and a
metadataSchemaon a strategy can designate individual keys asSTRICTLY_CONSISTENTrather than LLM-inferred. - Pass-through guarantee: When an application submits a short-term memory event with a value for a strictly consistent key, that exact value is preserved on the resulting long-term memory record without any LLM transformation or reinterpretation.
- Event grouping isolation: Events are grouped for extraction and consolidation based on their strictly consistent metadata values — events sharing the same values are processed together, while events with different values are always processed independently.
- No cross-value merging: Long-term memory records carrying different strictly consistent metadata values are never merged during consolidation, even if their semantic content is highly similar.
- Coexistence with LLM-inferred keys: A single memory resource can have both strictly consistent keys and LLM-inferred metadata keys operating simultaneously on the same strategy.
- Constraints: Up to three strictly consistent keys can be configured per strategy; keys must be of type STRING and must be declared in the memory's indexed keys (up to 10 indexed keys total per memory).
- Strategy support: The feature works across semantic, user preference, and episodic memory strategies, including custom strategy overrides.
- Querying: Strictly consistent metadata values can be used in filter expressions on
RetrieveMemoryRecordsorListMemoryRecords, combining up to five filters per query with AND logic.
Why It's Important
- Deterministic data integrity: Applications that require exact, auditable metadata values on memory records — such as compliance systems or regulated workflows — can now guarantee those values without relying on probabilistic LLM inference.
- Multi-tenant isolation: Each tenant's memory events can be tagged with a tenant identifier that strictly partitions extraction, consolidation, and retrieval, preventing any cross-tenant data leakage at the memory layer.
- Compliance boundary enforcement: Regulated records (e.g., HIPAA, GDPR-scoped data) can be tagged and isolated from standard records, ensuring they are never inadvertently merged with non-regulated content.
- Predictable retrieval scoping: Business dimensions like department, channel, or priority can be enforced as hard retrieval boundaries rather than soft semantic suggestions, making agent behavior more predictable and auditable.
- Reduced LLM error surface: Removing LLM inference from critical metadata fields eliminates a class of subtle bugs where the model misclassifies or inconsistently assigns metadata values across similar events.
- Operational simplicity: Developers can reason about memory partitioning using familiar application-layer concepts (e.g., tenant IDs, department codes) rather than designing complex prompt engineering to coerce consistent LLM outputs.
How It's Different
- LLM-inferred vs. strictly consistent: Previously, all metadata values were derived by the LLM during extraction, meaning values could vary based on phrasing, context, or model behavior; strictly consistent keys bypass this entirely.
- Grouping semantics: LLM-inferred metadata does not influence how events are grouped for consolidation, whereas strictly consistent keys act as hard partitioning boundaries that control the entire extraction and consolidation pipeline.
- Merge prevention: Standard semantic consolidation can merge records that are topically similar regardless of metadata; strictly consistent keys introduce a hard constraint that prevents cross-value merges unconditionally.
- Application-layer control: Strictly consistent metadata shifts control of critical classification decisions from the model to the application, aligning with software engineering patterns where business-critical identifiers are set programmatically.
- Namespace complement: Namespaces already isolate memories by primary entity; strictly consistent metadata adds a second, orthogonal dimension of isolation within a namespace without requiring separate memory resources.
When to Prefer It
- Multi-tenant SaaS platforms: When building agents that serve multiple customers from a shared memory resource and each tenant's data must be strictly isolated during both processing and retrieval.
- Regulated industries: When memory records must respect compliance boundaries (e.g., separating PHI from non-PHI, or GDPR-regulated from non-regulated records) and those boundaries cannot be left to LLM discretion.
- Department or role-scoped agents: When an enterprise agent must retrieve only records relevant to a specific department (e.g., Finance vs. Engineering) and cross-department contamination is unacceptable.
- Audit and traceability requirements: When downstream systems require a guaranteed, immutable metadata trail on memory records for audit logging, lineage tracking, or forensic review.
- High-volume consolidation pipelines: When semantic similarity between records from different logical partitions could cause incorrect merges, and strict partitioning is needed to maintain data fidelity at scale.
- Hybrid metadata strategies: When some fields (e.g., sentiment, topic) can tolerate LLM inference variability while others (e.g., tenant ID, regulatory classification) must be exact — both can coexist on the same memory resource.
Availability
- GA status: Generally available as of June 15, 2026.
- Regional availability: Available in all AWS Regions where Amazon Bedrock AgentCore Memory is supported.
- Supported strategies: Semantic, user preference, and episodic memory strategies, including custom overrides.
- Key limits: Up to three strictly consistent keys per strategy; up to 10 indexed keys total per memory resource.
- Type constraint: Strictly consistent keys must be of type STRING and must be declared as indexed keys on the memory resource before use.
- Indexed key immutability: Once an indexed key is added to a memory resource, it cannot be removed, so key naming should be planned carefully.
- Pricing: No separate pricing tier announced; subject to standard Amazon Bedrock AgentCore Memory pricing.
- Documentation: Available at the Amazon Bedrock AgentCore Developer Guide under "Structured metadata for long-term memories."
