Save and retrieve memories with Context Engine

Memory enables an agent to save useful task context in an AI index and retrieve it in later interactions. This helps agents reuse findings, decisions, and effective approaches without deriving them again.

Memories are shared with everyone who can access the AI index. They are not private to a user or conversation. Do not store personal, sensitive, or user-specific information in memory.

This guide demonstrates the memory workflow with an Elastic Agent Builder agent. Compatible external agents can use the same memory operations through tools exposed over Model Context Protocol (MCP).

You need:

• An Elastic Stack 9.6 deployment with an Enterprise license, or an Elastic Cloud Serverless project.
• Context Engine enabled in the current Kibana space.
• Permission to change global Advanced Settings.
• The All privilege for the Context Engine feature.
• Permission to create or edit a custom Elastic Agent Builder agent.
• The read, write, and view_index_metadata privileges for the AI index's backing Elasticsearch index.

Memory must be enabled globally before you can enable it for an AI index:

1. In Kibana, open Stack Management → Advanced Settings.
2. Select the Global Settings tab.
3. Search for Context Engine memory.
4. Turn on Context Engine memory (contextEngine:memoryEnabled).

This setting applies to all users and spaces in the current Elastic deployment or Serverless project. It makes the memory controls and tools available in spaces where Context Engine is enabled, but it does not enable memory for every AI index.

You can enable memory when you create an AI index or on an existing user-created AI index:

1. Open Context from the Kibana navigation.
2. Create an AI index, or open an existing AI index that you can edit.
3. In Memory, confirm that Enable memory is on. When the global memory setting is enabled, memory is on by default for user-created AI indices, including existing ones.
4. Confirm that the AI index description identifies the task context that belongs in the index and the questions it should support.

An AI index can contain both automation-generated Knowledge Indicators (KIs) and agent-written memories. You cannot enable memory for a managed AI index in Kibana. Memory writes are not supported for an AI index whose destination is an index pattern.

Assign the memory-enabled AI index to a custom agent:

1. In Elastic Agent Builder, open Manage components, then select Agents.
2. Create an agent, or open an agent that you can edit.
3. For an existing agent, select Edit agent settings. For a new agent, remain on the Settings tab.
4. In AI Indices, select the memory-enabled AI index under Additional indices.
5. Save the agent, then start a new conversation with it.

With Context Engine enabled, Elastic Agent Builder automatically includes the ki-retrieval skill for chat agents. When the agent loads the skill, it receives the following retrieval tools:

• platform.context_engine.list_ai_indices
• platform.context_engine.describe_ai_index
• platform.context_engine.query_ai_indices

Because the assigned AI index has memory enabled, Elastic Agent Builder also gives the agent the following memory tools:

• platform.context_engine.remember
• platform.context_engine.forget

You do not need to add the skill or tools manually.

Ask the agent to save a concrete, reusable fact. For example:

Remember this for future billing investigations: After changing the tax configuration, refresh the `orders-enriched` transform before validating invoice totals.
		

Confirm that the agent calls platform.context_engine.remember and selects the intended AI index. The tool returns the ID of the stored memory.

To inspect the memory in Kibana:

1. Open Context, then select the AI index.
2. Open the Knowledge Indicators tab.
3. Filter the list to the memory.session_fact type.
4. Confirm that the title, description, and content identify the saved fact.

You can also select View raw docs in Discover to inspect the stored document in Discover.

Test cross-conversation retrieval so the answer cannot come from the original conversation history:

1. Start a new conversation with the same agent.

2. Ask a question that depends on the saved memory without repeating its content. For example:

   What should I do before validating invoice totals after a tax configuration change?
   		

3. Confirm that the agent queries the intended AI index and uses the stored memory in its answer.

The agent recalls active, unexpired memories through the same AI index query operation used to retrieve KIs.

Ask the agent to forget a memory when its information is no longer valid or useful:

1. In a conversation with the agent, identify the memory and ask the agent to forget it. For example:

   Forget the memory about refreshing the `orders-enriched` transform.
   		

2. Confirm that the agent retrieves the correct memory and calls platform.context_engine.forget with its ID.

3. Start another conversation and repeat the recall question.

4. Confirm that the forgotten memory is not returned.

Forgetting a memory marks it as deleted so standard memory retrieval excludes it. It does not remove the stored document from Elasticsearch.

Memory has the following access and retention behavior:

• A memory belongs to its AI index, not to the user or conversation that created it. Anyone who can access the AI index can retrieve the memory.
• The AI index registry and agent assignment are scoped to a Kibana space. The Context Engine tools discover and query AI indices registered in the current space.
• Space access does not grant access to an AI index's backing Elasticsearch index. Reading and writing memories also requires the corresponding Elasticsearch index privileges.
• memory.session_fact records a granular fact learned during a task. memory.session records a broader synthesis of what was tried, what worked, and what to do differently.
• By default, a memory expires 90 days after it is last written. The remember tool can set another expiration time or create a non-expiring memory. Expiration excludes a memory from standard retrieval but does not delete its stored documents.
• The remember tool can revise a memory by writing the same memory ID. It does not automatically detect or merge duplicate memories.
• The forget tool creates a deletion tombstone. Forgotten memories remain stored but are excluded from retrieval.
• Disabling memory for an AI index prevents agents from writing, revising, or forgetting memories in that index, but it does not delete stored memories. If none of an agent's assigned AI indices have memory enabled, Elastic Agent Builder does not add the memory tools.
• Agent traces do not create memories automatically.

The following issues are specific to memory setup and retrieval:

For shared list, describe, and query failures, refer to Retrieve context from an AI index.