Memory

Memory lets agents store and recall information across conversations. Agents can persist facts, preferences, context, and instructions that survive between runs, making them more useful over time.

Memories page

In the UI

Navigate to Agents in the sidebar and open an agent’s detail page
Click the Memory tab to view all memories associated with this agent
Click Add Memory to create a memory manually
Enter the content, select a type and scope, and click Save
To edit or delete a memory, click the row and use the action buttons

Agents can also create memories automatically during conversations when configured with memory-enabled instructions.

Via the API

Create a memory


curl -X POST /memories \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "content": "The user prefers concise responses under 200 words",
    "type": "preference",
    "scope": "agent",
    "agentId": "agent-uuid"
  }'

List memories


curl "/memories?agentId=agent-uuid" \
  -H "Authorization: Bearer $TOKEN"

Update a memory


curl -X PATCH /memories/{id} \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "content": "The user prefers concise responses under 150 words"
  }'

Delete a memory


curl -X DELETE /memories/{id} \
  -H "Authorization: Bearer $TOKEN"

Search memories

Memories are embedded as vectors for semantic search. Query by meaning, not just keywords:


curl -X POST /memories/search \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "response length preference",
    "agentId": "agent-uuid",
    "limit": 5
  }'

Memory types

Type	Description	Example
`fact`	Objective information the agent has learned	”The company uses PostgreSQL 16”
`preference`	User or behavioral preferences	”Prefers bullet points over paragraphs”
`context`	Background context for ongoing work	”Currently migrating from v2 to v3 API”
`episode`	Record of a past interaction or event	”On 2026-03-15, user asked to reset their API keys”
`instruction`	Standing instructions that guide behavior	”Always include error handling in code examples”

Scopes

Scope	Visibility	Use case
`agent`	Single agent only	Agent-specific knowledge and preferences
`shared`	All agents in the organization	Cross-agent knowledge (e.g., company coding standards)
`global`	All agents across all organizations (admin)	Platform-wide instructions

Configuration reference

Field	Type	Description
`content`	string	The memory text
`type`	string	One of: `fact`, `preference`, `context`, `episode`, `instruction`
`scope`	string	One of: `agent`, `shared`, `global`
`agentId`	string	Required when scope is `agent`
`metadata`	object	Optional key-value pairs for tagging

How agents use memory

During execution, the agent’s memory module:

Retrieves relevant memories via semantic search using the current conversation context
Injects matching memories into the LLM prompt as additional context
Optionally auto-saves new facts discovered during the conversation

Configure memory behavior in the agent’s settings under Memory — toggle auto-save, set the maximum number of memories injected per turn, and choose which scopes to include.