Amazon Bedrock AgentCore Memory Guide: Strategies, Namespaces, and Retrieval

InYourLeague Research · 2026-08-29

Answer in brief

Amazon Bedrock AgentCore Memory combines session-scoped short-term context with strategy-driven long-term records organized by namespace. Retrieval can combine semantic search with structured metadata filters, while failed ingestions should follow the documented redrive path. The official source does not publish a selectable model ID for this product.

Key facts at a glance

Failure modes and verification

FAQ

What is the difference between short-term and long-term memory?

Short-term memory preserves turn-by-turn context within one session. Long-term memory extracts and stores preferences, important facts, and summaries across sessions, as described in the Amazon Bedrock AgentCore Memory overview.

How should actors and sessions be isolated?

Use a session as the short-term conversation boundary and a namespace as the persistent boundary for a primary entity such as a user or tenant. The metadata guide explains namespace-level entity isolation.

Is semantic search a memory strategy?

The supplied sources describe built-in, built-in with overrides, and custom strategy configuration. Semantic search is performed through RetrieveMemoryRecords, so the evidence supports treating it as retrieval behavior rather than a named strategy; see the long-term memory guide.

When can structured metadata be used in filters?

A metadata key must be declared as an indexed key before it can be referenced by metadataFilters. The structured metadata guide also states that adding an indexed key does not backfill existing records.

How are failed ingestions redriven, and which model ID should be selected?

Use the “Redrive failed ingestions” procedure in the official long-term memory guide; the supplied excerpt does not specify commands or retry limits. The official source does not publish a selectable model ID for this product.

Sources and freshness

• Official source
• Official source
• Official source
• Last verified: 2026-08-29

Extended guide

Use Amazon Bedrock AgentCore Memory with two explicit boundaries: keep immediate conversation history inside a session, and place persistent insights in a namespace associated with the relevant actor or primary entity. Configure a memory strategy before expecting long-term extraction, then retrieve records with semantic search and, where required, metadata filters.

Operating model

Implementation sequence

1. Create and configure memory. Create the memory resource and add one or more strategies to enable long-term memory. The long-term memory guide identifies built-in, built-in with overrides, and custom strategy configuration.

2. Define isolation boundaries. Use the session boundary for short-term exchanges. Use namespaces for persistent actor or entity separation. The supplied official excerpts do not specify actor or session identifier field names, so applications should not infer undocumented identifiers from this guide.

3. Declare structured metadata. Add indexed keys with CreateMemory or UpdateMemory; only indexed keys can appear in metadataFilters. Amazon Bedrock AgentCore Memory supports STRING, STRINGLIST, and NUMBER indexed values. A memory can declare up to 10 indexed keys. An indexed key cannot be removed after addition, and adding one does not backfill older records.

4. Configure extraction. A strategy can define memoryRecordSchema.metadataSchema so the LLM extracts metadata while producing memory records. Validation can normalize allowed values. For application-known classifiers such as department or agent identity, STRICTLY_CONSISTENT copies supplied values unchanged instead of asking the LLM to infer them.

5. Ingest and retrieve. Supply content through CreateEvent or IngestData, or create records with BatchCreateMemoryRecords. Use RetrieveMemoryRecords for semantic search with optional pre-filtering. Use ListMemoryRecords for metadata-only filtering. Queries can combine up to 5 filters with AND logic.

6. Handle ingestion failures. The official long-term guide provides a dedicated “Redrive failed ingestions” section. Follow that procedure rather than inventing retry commands or limits; the supplied excerpts do not specify those operational details.

Release checklist

• Short-term context is scoped to one session.
• Long-term records use actor- or entity-aligned namespaces.
• Every filterable metadata key is indexed.
• Controlled values use validation or STRICTLY_CONSISTENT where appropriate.
• Semantic and metadata-filtered retrieval are tested separately.
• Failed-ingestion redrive is covered by operations procedures.

Verification boundary

Verified against the supplied official documentation on 2026-08-29. The official source does not publish a selectable model ID for this product.

Model availability note: The official source does not specify a selectable model ID.