# User Memory
Let your AI remember facts about users across conversations for personalized experiences.
---
User Memory enables your AI to remember important facts about users across conversations. Instead of starting fresh every time, your AI can recall preferences, past interactions, and relevant details.
## What Is User Memory?
User Memory automatically extracts and stores durable facts from conversations:
| Memory Type | Examples |
|-------------|----------|
| **Fact** | "User's company is Acme Corp", "Works as a CTO" |
| **Preference** | "Likes concise answers", "Prefers email over phone" |
| **Instruction** | "Always reply in Spanish", "Don't use bullet points" |
| **Context** | "Evaluating the enterprise plan", "Migrating off a legacy CRM" |
Each memory is also stored in one of two layers: **Fact** (kept the longest) or **Episodic** (for things that matter for a while).
### How Long Memories Last
Every memory gets its own lifetime when it is saved, based on what it says:
| Lifetime | Length | Typical memories |
|----------|--------|------------------|
| **Short-term** | 14 days | "Has a flight tomorrow", "Busy this week" |
| **Situational** | 90 days | "Currently working on a remodel", episodic notes, context |
| **Standard** | 1 year | Most facts and preferences |
| **Durable** | 2 years | Names, allergies, standing instructions |
When a person says the same thing again, the memory's lifetime starts over from that day. When they correct it ("I moved to Braga"), the new version replaces the old one.
When a lifetime ends, your agent stops using the memory at once, and its text is permanently deleted soon after. Chipp keeps only a record that a memory existed and when it expired, never what it said.
## How It Works
```
User: "I'm the CTO at Acme Corp and we're evaluating your enterprise plan"
↓
[AI replies]
↓
[Memory extraction runs in the background]
↓
Memories stored:
- Role: CTO
- Company: Acme Corp
- Interest: Enterprise plan
↓
[Next conversation]
↓
AI: "Welcome back! Last time we discussed enterprise pricing for Acme Corp..."
```
## Enabling User Memory
Memory is on by default for new apps. To review or change it:
**1.**
Open Memory Settings
In your app's builder, open **Brain** and then the **Capabilities** tab. Find the **Memory** card.
**2.**
Auto-extract memories
Toggle on **Auto-extract memories**. Your AI then learns preferences, facts, and context from conversations with each consumer.
**3.**
Recall memories in new conversations
Toggle on **Recall memories in new conversations** to carry saved facts about a consumer into their later conversations, and to let your AI search them and get a summary. Turn this off for assessment, intake, or eligibility flows, where each conversation must be judged only on what the person says in it.
**4.**
List saved memories
Toggle on **List saved memories** to let your AI list every saved memory about the consumer it is talking to, page by page, including replaced and deleted versions. Turn it off if the full list must never appear in chat or in channels like Slack.
**5.**
Save and change memories
Toggle on **Save and change memories** to let your AI save new memories, and update or delete them when the consumer asks. With recall and listing off, your AI can still save, but it only sees what it saved in the current conversation, never other stored memories.
**6.**
Consumer memory access (Optional)
Toggle on **Consumer memory access** to let consumers view and delete their own memories from the chat menu. This is recommended for GDPR compliance.
Turn all four switches off to turn memory off completely. The three AI switches (recall, list, save and change) apply the same way in web chat and in every channel, such as Slack.
If you set up your app before these three switches existed, nothing changes: an app with recall off still has listing and saving off, and an app with recall on keeps all three on until you change them.
To steer what the AI remembers or ignores, put that guidance in your system prompt (see "Memory in Your System Prompt" below).
## Memory Categories
### Fact
Personal facts about the user:
- Name, job, and location
- Company information
- Dietary restrictions, family, pets
### Preference
Likes, dislikes, and style:
- Communication style (formal/casual)
- Response length preference
- Preferred formats (bullet points, paragraphs)
### Instruction
How the user wants to be treated or how they want tasks done:
- "Always answer in Spanish"
- "Show the code before the explanation"
### Context
Ongoing projects, goals, or situations:
- Problems they're trying to solve
- Products they're evaluating
- Deadlines they mentioned
## Viewing Stored Memories
Go to **Insights > Memory** in the builder and select a consumer to see:
- All memories for that consumer
- When each memory was created
- Each memory's lifetime, when it expires, and when the person last confirmed it
Select **Show expired memories** to see the memories that reached the end of their lifetime. Their text is gone; only the dates and the lifetime remain.
Filter the list by memory type (Preference, Fact, Instruction, Context, Experiential) or by layer (Episodic or Fact).
## Managing Memories
### Deleting Memories
Remove memories that are:
- Incorrect
- Outdated
- No longer relevant
Click the delete icon on a memory and confirm.
### Clearing All Memories for a Consumer
For compliance or cleanup, select the consumer in **Insights > Memory** and click **Clear all**.
### Reviewing and Purging Memories as an Admin
Through the Chipp MCP, an admin can review and remove a consumer's memories even when your AI's memory switches are off. Those switches only control what your AI can do.
- `list_user_memories` pages through every memory of one consumer, with each memory's ID and status. Add `include_history: true` to also see replaced and deleted versions. You need write access to **App settings** for the app's workspace.
- `purge_user_memory_chain` permanently deletes one memory and every earlier and later version of it. It needs `confirm: true` and write access to **Danger zone** (a workspace owner, or an organization owner or admin). Each purge is recorded in your organization's audit log, without the memory text. It cannot be undone.
See the [MCP tools reference](/docs/guides/mcp/tools-reference#list_user_memories) for the parameters.
## Best Practices
### What to Remember
**Do remember:**
- Professional context (company, role, industry)
- Stated preferences
- Key topics discussed
- Problems they're solving
**Don't remember:**
- Sensitive personal information
- Temporary states ("having a bad day")
- Information they asked you to forget
- Confidential business data without consent
### Privacy Considerations
> **Warning:** Always inform users that your AI has memory capabilities. Consider adding to your system prompt:
>
> "I can remember relevant details from our conversations to provide better assistance. Let me know if you'd prefer I not remember something."
### Memory in Your System Prompt
Help your AI use memories effectively:
```
You have access to memories about this user from past conversations.
Use these to:
- Personalize greetings and responses
- Avoid asking questions they've already answered
- Reference their specific context
Don't:
- Mention that you "remember" things in a creepy way
- Assume memories are 100% accurate
- Reference memories if they seem outdated
```
## How Memory Extraction Works
Memory extraction happens automatically, in the background, after each AI reply:
1. **Turn completes**: The AI finishes a reply (stopped or aborted replies are skipped)
2. **AI analysis**: A Claude Haiku model reads the exchange for memorable facts
3. **Structured extraction**: Facts are given a type (fact, preference, instruction, context) and a layer (fact or episodic) and a lifetime
4. **Storage**: Memories are stored against the consumer
Extraction is skipped for apps with Zero Data Retention turned on.
## Use Cases
### Customer Support
Remember:
- Product version they're using
- Past issues they've reported
- Preferred contact method
- Technical expertise level
### Sales Conversations
Remember:
- Company size and industry
- Budget discussions
- Decision timeline
- Stakeholders involved
### Personal Assistants
Remember:
- Work schedule preferences
- Project names and contexts
- Communication preferences
- Recurring tasks
### Educational Apps
Remember:
- Learning progress
- Topics mastered
- Areas needing work
- Preferred learning style
## Troubleshooting
### Memories Not Being Extracted
1. Verify **Auto-extract memories** is on in the Memory card
2. Check that Zero Data Retention is off for the app
3. Allow a few moments after a reply for extraction to finish
4. Verify the user is identified (not anonymous)
### Incorrect Memories
1. Delete incorrect memories
2. Add guidance to your system prompt about what to remember and what to ignore
### Too Many Memories
1. Add system prompt guidance to prioritize durable facts
2. Periodically clean up outdated memories
## Limitations
Current limitations to be aware of:
- **Extraction delay**: Memories from a reply are written in the background, so they may not be usable in the very next message
- **Anonymous users**: Memories need an identified consumer
- **Token limits**: Very long conversations may be truncated for extraction
## Using with MCP
You can access user memories programmatically using the Chipp MCP Server.
### Available Tools
| Tool | Description |
|------|-------------|
| `get_user_memories` | Retrieve memories stored for a specific consumer (`app_id`, `consumer_id`, optional type filter and limit) |
| `list_user_memories` | Admin inventory of one consumer's memories, paged, with optional history (`include_history`) |
| `purge_user_memory_chain` | Permanently delete one memory and its full version history (`confirm: true`, Danger zone access) |
### Example Prompts
```
"Show me the memories my support bot has for consumer 8f2c..."
"What facts has my app learned about this user?"
"List only the instruction memories for this consumer"
```
### Workflow: Reviewing User Insights
```
You: Show me the memories my sales bot has for john@acme.com
AI: [Looks up the consumer, then calls get_user_memories]
Found 6 memories:
- fact: CTO at Acme Corp, 500+ employees
- context: Evaluating the enterprise plan, budget approved
- preference: Prefers short, direct answers
...
```
> **Note:** MCP server access requires a Studio plan or higher. See the [MCP Setup Guide](/docs/guides/mcp/setup) to get started.
> **Note:** **Individual vs. group memory.** User Memory stores facts about a single user -- preferences, past interactions, and extracted insights. [Circles](/docs/guides/circles) complement this with shared group memory: a persistent context space that multiple users and AI agents read and write together. Use both for rich, personalized, and collaborative AI experiences.