Memory

Source: src/AWS/BedrockAgentCore/Memory.ts

An Amazon Bedrock AgentCore Memory — managed short- and long-term memory for AI agents.

Short-term memory stores raw session events (turn-by-turn conversation); optional memoryStrategies asynchronously extract long-term records (semantic facts, summaries, user preferences) into queryable namespaces.

Provisioning is asynchronous: the provider waits for the memory to reach ACTIVE (~2-3 minutes) before returning.

Creating Memories

Short-Term Memory Only

import * as AgentCore from "alchemy/AWS/BedrockAgentCore";

const memory = yield* AgentCore.Memory("SessionMemory", {
  eventExpiryDuration: "30 days",
});

Memory with a Semantic Long-Term Strategy

const memory = yield* AgentCore.Memory("AgentMemory", {
  eventExpiryDuration: "90 days",
  memoryStrategies: [
    {
      semanticMemoryStrategy: {
        name: "facts",
        namespaces: ["facts/{actorId}"],
      },
    },
  ],
});

Using Memory from a Function

// init
const createEvent = yield* AgentCore.CreateEvent(memory);
const listEvents = yield* AgentCore.ListEvents(memory);

return {
  fetch: Effect.gen(function* () {
    // runtime
    yield* createEvent({
      actorId: "user-1",
      sessionId: "session-1",
      eventTimestamp: new Date(),
      payload: [
        {
          conversational: {
            role: "USER",
            content: { text: "My favorite color is teal." },
          },
        },
      ],
    });
    const events = yield* listEvents({
      actorId: "user-1",
      sessionId: "session-1",
    });
    return HttpServerResponse.json({ count: events.events.length });
  }),
};