AWS AgentCore Long-Term Memory

[mathieu-stennier]

Summary

Implement an outbound Camunda connector for AWS AgentCore Long-Term Memory, enabling BPMN processes and AI agents to retrieve persistent knowledge — user preferences, semantic facts, and interaction summaries — that survives across sessions and process instances.

Motivation

While short-term memory handles within-session context, long-term memory gives agents the ability to learn and personalize over time. AgentCore Long-Term Memory automatically extracts insights from raw conversation events (facts, preferences, summaries) and makes them retrievable via semantic search. In a Camunda context, this bridges the gap between individual process executions: an agent handling a returning customer can recall their preferences, prior issues, and communication style without re-asking. This is the foundation for truly personalized, context-aware business process automation.

Use Cases in Camunda

• Customer relationship continuity: An insurance claims agent recalls that the customer prefers email communication, has a history of flood claims, and previously expressed frustration with response times — all extracted automatically from prior process interactions.
• Agent learning across process instances: A support agent remembers that a particular workaround resolved a recurring issue for a customer, avoiding redundant troubleshooting in future tickets.
• Cross-agent knowledge sharing: Multiple agents operating in different processes (sales, support, billing) share a memory namespace, building a unified customer profile.
• Compliance-aware memory: Configure extraction rules and expiry policies to ensure the agent only retains what's appropriate, with full auditability.

AWS Setup (High-Level)

1. Open the Amazon Bedrock AgentCore console → navigate to Memory.
2. Create a memory resource with long-term memory enabled — this provisions both short-term event storage and the long-term extraction pipeline.
3. Configure extraction strategy: use MANAGED (AWS-managed extraction of facts, preferences, summaries) or SELF_MANAGED (bring your own extraction pipeline).
4. Set consolidation and retention policies as needed.
5. Create an IAM policy granting permissions for bedrock-agentcore:RetrieveMemoryRecords, bedrock-agentcore:WriteMemoryEvents, and related long-term memory actions.
6. Note the Memory ID and region.

Connector Specification

Operations

Operation	Description
Retrieve Memory Records	Semantic search across long-term memory. Returns relevant facts, preferences, and summaries matching a query.
List Memory Records	List stored memory records for a given namespace/actor, with optional filtering.

Key Input Fields

• Memory Resource ID (required)
• Query text (required for Retrieve — supports FEEL expressions)
• Namespace (required — scopes memory, e.g. per customer or per process definition)
• Actor ID (optional — identifies the user or agent whose memory to query)
• Max results (optional)
• AWS Region (required)
• Authentication — reuse existing AWS auth pattern

Output

• Memory records with content, type (fact / preference / summary), confidence score
• Record metadata (creation time, source session, extraction method)

References

• AgentCore Memory — long-term walkthrough
• AgentCore Memory blog post
• AgentCore pricing — Memory
• Camunda + AWS agentic orchestration