Extending conversational memory in Kiro CLI using Amazon Bedrock AgentCore Memory

TL;DR

• Implementing a custom Model Context Protocol (MCP) server enables Kiro CLI to maintain conversational context across multiple sessions using Amazon Bedrock AgentCore Memory.
• The solution utilizes a two-stage retrieval strategy, first attempting semantic search via the retrieve_memory_records API and falling back to direct event-level content matching.
• The MCP server exposes three distinct categories of tools: conversation management (search, store, retrieve), monitoring (usage statistics, config), and management (deletion of sessions/data).
• Actors can be identified by User ID (for personal use) or Project ID (for team isolation), utilizing a specific namespace strategy: /strategy/semanticMemoryStrategy/actor/{actorId}/session/{sessionId}.
• This integration allows agentic IDEs to recall preferences and insights from previous interactions, reducing the need for repetitive context-setting by developers.

Why It Matters

Agentic IDEs currently lack persistent memory, forcing developers to repeatedly provide context for complex, long-term projects, which significantly impacts productivity. By leveraging fully managed services like Amazon Bedrock AgentCore Memory through the MCP standard, developers can create context-aware workflows that learn from past interactions and retain project-specific details automatically. This approach bridges the gap between stateless AI interactions and continuous, personalized developer assistance.

Key Data

• Namespace Strategy: The solution uses the specific path structure /strategy/semanticMemoryStrategy/actor/{actorId}/session/{sessionId} for memory organization.
• Retrieval Method: A two-stage process is used: 1) Semantic search using retrieve_memory_records API, and 2) Fallback to direct event-level content matching scanning raw payloads.
• Actor Identification Options: Two types are supported: User ID (derived from the USER environment variable) and Project ID (user-defined string), with User ID being the recommended default for personal use.
• Protocol Connection: Kiro CLI connects to the custom MCP server specifically through STDIO protocols.

Technical Details

• Architecture Components: The system consists of three main parts: Amazon Bedrock AgentCore Memory (for persistent storage with semantic search), a Custom MCP Server (bridging memory operations to MCP-compatible clients), and Kiro CLI (the client interface).
• Tool Categories: The MCP server provides tools grouped into three functions: Conversation tools (search by topic/timeframe, store with session tracking, retrieve content), Monitoring tools (view memory usage and configs), and Management tools (delete specific sessions or data).
• Implementation Steps: Setup requires cloning the repository, creating a Python virtual environment to install dependencies, and running setup_bedrock_agentcore_memory.py to create the AWS resource and generate the Kiro agent configuration.
• Configuration Files: The agent configuration is stored in agent/kiro_memory.json, and shell hooks are copied to ~/.kiro/hooks/ to manage prompt caching, preference loading, and conversation storage.
• Prerequisites: Users need access to an AWS account, IAM permissions following the principle of least privilege, an AWS Builder ID, and a configured Kiro CLI installation.

Industry Insight

• The standardization of MCP allows third-party memory services to integrate seamlessly with emerging agentic IDEs, suggesting a future where "memory" is a pluggable service rather than a built-in feature of the AI model.
• Supporting both User ID and Project ID actor strategies highlights the importance of data isolation in enterprise environments, allowing teams to share project memory while keeping individual developer preferences separate.
• The fallback mechanism to raw content matching addresses a critical reliability issue in semantic search systems, ensuring that developers can always retrieve context even if the vector search indexing has not yet completed.

FAQ

Q: What protocol does Kiro CLI use to communicate with the custom MCP server?
A: Kiro CLI connects to the MCP server through STDIO protocols to store and retrieve conversational history.

Q: How does the system handle cases where semantic search has not finished processing a conversation?
A: The solution uses a two-stage retrieval strategy; if the semantic search via retrieve_memory_records is not complete, it falls back to direct event-level content matching by scanning raw conversation payloads.

Q: Which actor identifier is recommended for personal versus team usage?
A: User ID is recommended for personal use, while Project ID is recommended for team or project-based memory isolation.