Overview
Agent Memory allows your agent to detect, store, and recall user-specific information across conversations. When enabled, the agent analyzes messages for useful context, stores it as memory records, and injects approved memories into future interactions — creating a personalized experience for each user. Agent Memory is useful for remembering:- User preferences — preferred tone, response format, or workflow style.
- Stable user facts — role, team, company, or recurring needs.
- Important context — information that helps personalize future interactions.
How Agent Memory Works
The agent analyzes conversations
User messages are automatically analyzed in the background for memorable preferences, facts, or context. The agent performs a duplicate check against existing memories and assigns a confidence score between 0.0 and 1.0.
Candidates are stored as pending
Detected memory candidates are saved with a
pending status. Pending memories expire after 7 days if not reviewed.Memories are approved or rejected
Admins review pending memories and approve or reject them through the Memory Management panel or via the REST API.
Memory Types
The agent can capture and store three types of memory:| Type | Description | Examples |
|---|---|---|
Preference | User preferences such as tone, format, or workflow style. | ”User prefers bullet point summaries”, “User likes concise technical answers” |
Fact | Stable facts about the user such as role, team, or company. | ”User is CTO at StartupCorp”, “User specializes in Python and machine learning” |
Context | Situational context that helps personalize future interactions. | ”User has a product launch deadline next month”, “User’s team uses Agile methodology” |
Memory Statuses
| Status | Description |
|---|---|
pending | Memory has been detected and is awaiting review. |
approved | Memory has been approved and is actively used in agent responses. |
rejected | Memory was declined and will not be used. |
Memory Lifecycle
Creation
- User sends a message.
- Agent analyzes the message in the background.
- AI detects memorable information and performs a duplicate check.
- Memory is created with
pendingstatus and a confidence score assigned. - A 7-day expiration timer starts.
Approval
- Admin reviews the pending memory.
- Memory is approved or rejected.
- If approved, memory becomes active and is injected into future conversations.
- If rejected, the memory is archived.
Usage
- User starts a conversation.
- System loads approved memories for that user.
- Memories are formatted by type and added to the agent’s system prompt.
- Agent uses memories to personalize responses.
Expiration
- Pending memories expire after 7 days if not approved.
- Approved memories do not expire unless manually deleted.
- Expired pending memories are automatically rejected.
Examples
Example 1: Preference Memory
Example 1: Preference Memory
User message:
Future impact: The agent will default to Python code examples whenever the user asks for code.
“I prefer to receive code examples in Python rather than pseudocode.”Detected memory:
| Field | Value |
|---|---|
| Type | preference |
| Content | ”User prefers Python code examples over pseudocode.” |
| Confidence | 0.90 |
Example 2: Fact Memory
Example 2: Fact Memory
User message:
Future impact: The agent will understand the user’s role and company context in all future conversations.
“I’m the CTO at StartupCorp and we’re building a SaaS platform.”Detected memory:
| Field | Value |
|---|---|
| Type | fact |
| Content | ”User is CTO at StartupCorp building a SaaS platform.” |
| Confidence | 0.95 |
Example 3: Context Memory
Example 3: Context Memory
User message:
Future impact: The agent will prioritize speed and urgency in its responses.
“We’re launching our product next month, so I need everything ready quickly.”Detected memory:
| Field | Value |
|---|---|
| Type | context |
| Content | ”User has a product launch deadline next month requiring quick turnaround.” |
| Confidence | 0.85 |