Crear un agente con memory stores

Usa memory stores cuando necesites contexto durable que pueda usarse entre usuarios, sesiones y agentes, y al que puedan acceder aplicaciones externas. Los memory stores son útiles para almacenar notas de cuentas, manuales de soporte, informes de errores o cualquier contenido que quieras reutilizar en muchas ejecuciones de agentes. Referencia de API:

1. Crea el memory store

Crea un store para un cuerpo durable de conocimiento. Usa instructions para describir cómo deben escribirse y mantenerse las entradas.

mka1 llm memory-stores create --body '{
  "name": "Acme support memory",
  "description": "Durable support notes for Acme Corp.",
  "instructions": "Keep entries concise. Use one file per durable fact, customer preference, or open issue.",
  "visibility": "workspace",
  "metadata": {
    "account_id": "acct_acme"
  }
}' \
  -H 'X-On-Behalf-Of: <end-user-id>'

import { SDK } from "@meetkai/mka1";

const sdk = new SDK({
  bearerAuth: "Bearer <mka1-api-key>",
});

const store = await sdk.llm.memoryStores.create({
  xOnBehalfOf: "<end-user-id>",
  createMemoryStoreRequest: {
    name: "Acme support memory",
    description: "Durable support notes for Acme Corp.",
    instructions:
      "Keep entries concise. Use one file per durable fact, customer preference, or open issue.",
    visibility: "workspace",
    metadata: {
      account_id: "acct_acme",
    },
  },
});

console.log(store.id);

curl https://apigw.mka1.com/api/v1/llm/memory_stores \
  --request POST \
  --header 'Content-Type: application/json' \
  --header 'Authorization: Bearer <mka1-api-key>' \
  --header 'X-On-Behalf-Of: <end-user-id>' \
  --data '{
    "name": "Acme support memory",
    "description": "Durable support notes for Acme Corp.",
    "instructions": "Keep entries concise. Use one file per durable fact, customer preference, or open issue.",
    "visibility": "workspace",
    "metadata": {
      "account_id": "acct_acme"
    }
  }'

La respuesta incluye un ID como mem_store_.... visibility determina si este memory store puede usarlo solo un X-On-Behalf-Of específico (private) o si puede acceder cualquier usuario, sin importar el X-On-Behalf-Of (workspace). Consulta la referencia para crear un memory store.

2. Agrega entradas de memoria

Las entradas se direccionan por path. Usa rutas predecibles para que el agente encuentre archivos en el store montado.

mka1 llm memory-stores create-entry \
  --memory-store-id mem_store_123 \
  --body '{
    "path": "accounts/acme/preferences.md",
    "content": "# Acme preferences\n\n- Prefer concise weekly status summaries.\n- Escalate production incidents to the on-call channel before creating a ticket.",
    "metadata": {
      "kind": "customer-preferences"
    }
  }' \
  -H 'X-On-Behalf-Of: <end-user-id>'

const entry = await sdk.llm.memoryStores.createEntry({
  memoryStoreId: "mem_store_123",
  xOnBehalfOf: "<end-user-id>",
  createMemoryEntryRequest: {
    path: "accounts/acme/preferences.md",
    content:
      "# Acme preferences\n\n- Prefer concise weekly status summaries.\n- Escalate production incidents to the on-call channel before creating a ticket.",
    metadata: {
      kind: "customer-preferences",
    },
  },
});

console.log(entry.id, entry.contentHash);

curl https://apigw.mka1.com/api/v1/llm/memory_stores/mem_store_123/entries \
  --request POST \
  --header 'Content-Type: application/json' \
  --header 'Authorization: Bearer <mka1-api-key>' \
  --header 'X-On-Behalf-Of: <end-user-id>' \
  --data '{
    "path": "accounts/acme/preferences.md",
    "content": "# Acme preferences\n\n- Prefer concise weekly status summaries.\n- Escalate production incidents to the on-call channel before creating a ticket.",
    "metadata": {
      "kind": "customer-preferences"
    }
  }'

La respuesta incluye un ID de entrada y content_hash.

3. Crea el agente guardado

Crea un agente con una herramienta shell y monta el memory store en el entorno shell. Usa read_only para memoria de referencia. Usa read_write cuando el agente deba actualizar memoria escribiendo archivos en /mnt/memory/<label>. Cuando el agente escribe, edita o elimina archivos dentro de un montaje read_write, el sandbox sincroniza esos cambios de vuelta al memory store al terminar el comando shell.

mka1 agents create --body '{
  "name": "account-support-agent",
  "description": "Answers support questions with durable account context.",
  "model": "auto",
  "instructions": "Use the mounted Acme support memory for durable account context. Read files under /mnt/memory/acme_support before answering account-specific questions. When you learn a durable Acme fact, write or update the relevant markdown file under /mnt/memory/acme_support. Do not invent account facts that are not present in memory.",
  "tools": [
    {
      "type": "shell",
      "environment": {
        "type": "container_auto",
        "memory_stores": [
          {
            "store_id": "mem_store_123",
            "label": "acme_support",
            "access": "read_write",
            "description": "Durable support notes for Acme Corp.",
            "instructions": "Use these files as account context. Prefer specific files under accounts/acme/. Write durable account learnings back as concise markdown."
          }
        ]
      }
    }
  ],
  "tool_choice": "auto",
  "metadata": {
    "memory_store_id": "mem_store_123",
    "account_id": "acct_acme"
  }
}' \
  -H 'X-On-Behalf-Of: <end-user-id>'

const agent = await sdk.agents.createAgent({
  xOnBehalfOf: "<end-user-id>",
  createAgentRequest: {
    name: "account-support-agent",
    description: "Answers support questions with durable account context.",
    model: "auto",
    instructions:
      "Use the mounted Acme support memory for durable account context. Read files under /mnt/memory/acme_support before answering account-specific questions. When you learn a durable Acme fact, write or update the relevant markdown file under /mnt/memory/acme_support. Do not invent account facts that are not present in memory.",
    tools: [
      {
        type: "shell",
        environment: {
          type: "container_auto",
          memoryStores: [
            {
              storeId: "mem_store_123",
              label: "acme_support",
              access: "read_write",
              description: "Durable support notes for Acme Corp.",
              instructions:
                "Use these files as account context. Prefer specific files under accounts/acme/. Write durable account learnings back as concise markdown.",
            },
          ],
        },
      },
    ],
    toolChoice: "auto",
    metadata: {
      memory_store_id: "mem_store_123",
      account_id: "acct_acme",
    },
  },
});

console.log(agent.id);

curl https://apigw.mka1.com/api/v1/agents \
  --request POST \
  --header 'Content-Type: application/json' \
  --header 'Authorization: Bearer <mka1-api-key>' \
  --header 'X-On-Behalf-Of: <end-user-id>' \
  --data '{
    "name": "account-support-agent",
    "description": "Answers support questions with durable account context.",
    "model": "auto",
    "instructions": "Use the mounted Acme support memory for durable account context. Read files under /mnt/memory/acme_support before answering account-specific questions. When you learn a durable Acme fact, write or update the relevant markdown file under /mnt/memory/acme_support. Do not invent account facts that are not present in memory.",
    "tools": [
      {
        "type": "shell",
        "environment": {
          "type": "container_auto",
          "memory_stores": [
            {
              "store_id": "mem_store_123",
              "label": "acme_support",
              "access": "read_write",
              "description": "Durable support notes for Acme Corp.",
              "instructions": "Use these files as account context. Prefer specific files under accounts/acme/. Write durable account learnings back as concise markdown."
            }
          ]
        }
      }
    ],
    "tool_choice": "auto",
    "metadata": {
      "memory_store_id": "mem_store_123",
      "account_id": "acct_acme"
    }
  }'

Consulta la referencia para crear un agente.

4. Ejecuta el agente y actualiza memoria

El agente puede leer archivos montados con la herramienta shell. Si el store montado es read_write, también puede actualizar memoria escribiendo archivos bajo /mnt/memory/<label>. El sandbox restaura los archivos montados antes de ejecutar el comando, toma una instantánea y sincroniza los archivos cambiados al terminar.

mka1 agent-runs create \
  --agent-id agt_123 \
  --body '{
    "input": "Draft a support update for Acme about the production incident response plan. Use the mounted Acme memory before answering. Also remember that Acme now wants customer-facing incident summaries by 4 PM Central.",
    "metadata": {
      "memory_store_id": "mem_store_123",
      "source": "docs-recipe"
    }
  }' \
  -H 'X-On-Behalf-Of: <end-user-id>'

const run = await sdk.agentRuns.createAgentRun({
  agentId: "agt_123",
  xOnBehalfOf: "<end-user-id>",
  createAgentRunRequest: {
    input:
      "Draft a support update for Acme about the production incident response plan. Use the mounted Acme memory before answering. Also remember that Acme now wants customer-facing incident summaries by 4 PM Central.",
    metadata: {
      memory_store_id: "mem_store_123",
      source: "docs-recipe",
    },
  },
});

console.log(run.id);

curl https://apigw.mka1.com/api/v1/agents/agt_123/runs \
  --request POST \
  --header 'Content-Type: application/json' \
  --header 'Authorization: Bearer <mka1-api-key>' \
  --header 'X-On-Behalf-Of: <end-user-id>' \
  --data '{
    "input": "Draft a support update for Acme about the production incident response plan. Use the mounted Acme memory before answering. Also remember that Acme now wants customer-facing incident summaries by 4 PM Central.",
    "metadata": {
      "memory_store_id": "mem_store_123",
      "source": "docs-recipe"
    }
  }'

Usa recuperar una ejecución de agente para consultar el estado. Usa transmitir eventos de ejecución para mostrar progreso en vivo.

5. Inspecciona la memoria después

Después de que termine la ejecución, usa la Memory Stores API para revisar qué cambió. Los mismos archivos que el agente escribe en /mnt/memory/acme_support están disponibles como entradas de memoria.

mka1 llm memory-stores list-entries \
  --memory-store-id mem_store_123 \
  --limit 20 \
  -H 'X-On-Behalf-Of: <end-user-id>'

const entries = await sdk.llm.memoryStores.listEntries({
  memoryStoreId: "mem_store_123",
  limit: 20,
  xOnBehalfOf: "<end-user-id>",
});

console.log(entries.data.map((entry) => entry.path));

curl 'https://apigw.mka1.com/api/v1/llm/memory_stores/mem_store_123/entries?limit=20' \
  --header 'Authorization: Bearer <mka1-api-key>' \
  --header 'X-On-Behalf-Of: <end-user-id>'

Consulta la referencia para listar entradas de memoria.

Notas operativas

Usa montajes read_only para memoria de referencia y read_write para agentes que puedan editar memoria durable.
Deja que los agentes actualicen memoria escribiendo archivos en montajes read_write con la herramienta shell.
Usa actualizaciones directas de la Memory Stores API para semillas iniciales, revisiones, correcciones administrativas o importaciones.
Mantén las entradas pequeñas para que el agente inspeccione solo archivos relevantes durante la ejecución.

Comenzar

Respuestas

Funciones

Compute

CLI

Recetas

Comparativas

Infraestructura

Crear un agente con memory stores

1. Crea el memory store

2. Agrega entradas de memoria

3. Crea el agente guardado

4. Ejecuta el agente y actualiza memoria

5. Inspecciona la memoria después

Notas operativas

​1. Crea el memory store

​2. Agrega entradas de memoria

​3. Crea el agente guardado

​4. Ejecuta el agente y actualiza memoria

​5. Inspecciona la memoria después

​Notas operativas

1. Crea el memory store

2. Agrega entradas de memoria

3. Crea el agente guardado

4. Ejecuta el agente y actualiza memoria

5. Inspecciona la memoria después

Notas operativas