> ## Documentation Index
> Fetch the complete documentation index at: https://reedai-07fa30f1.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Memory and Context Issues

> Troubleshoot problems with AI memory, context retention, and conversation history in ZeroTwo

When the AI forgets previous interactions, provides inconsistent information, or fails to remember important details, you're experiencing memory or context issues. Learn how to identify and resolve these problems.

## Common memory problems

<AccordionGroup>
  <Accordion title="AI forgets earlier conversation">
    **Symptom**: AI doesn't remember details from earlier in the same conversation

    **Causes:**

    * Context window limit exceeded
    * Very long conversation history
    * Model limitations
    * Information not properly retained

    **Solutions:**

    1. **Summarize key points**: Explicitly remind the AI of important details
    2. **Use Memory feature**: Enable [Memory](/tools/memory) to retain facts across chats
    3. **Start new conversation**: Archive old chat and start fresh with summary
    4. **Reference specific messages**: Point to earlier responses
    5. **Reduce context**: Remove unnecessary messages

    ```text Effective reminder theme={null}
    "As we discussed earlier about the user authentication system with JWT tokens..."
    ```

    <Info>
      Most models have context windows between 8K-200K tokens. Long conversations consume this limit, causing older messages to be truncated.
    </Info>
  </Accordion>

  <Accordion title="Inconsistent information across chats">
    **Symptom**: AI provides different answers to similar questions in different conversations

    **Causes:**

    * Memory not enabled or not working
    * Different models used
    * Information not saved to Memory
    * Conflicting custom instructions

    **Solutions:**

    1. **Enable Memory**: Turn on [Memory feature](/tools/memory)
    2. **Verify memory content**: Check what's stored in Memory settings
    3. **Use same model**: Stick with one model for consistency
    4. **Set custom instructions**: Create [consistent instructions](/chat/custom-instructions)
    5. **Share context explicitly**: Don't rely solely on Memory

    <Warning>
      Memory is model-specific in some cases. Information stored while using Claude may not transfer to GPT-4 sessions.
    </Warning>
  </Accordion>

  <Accordion title="Memory contains outdated information">
    **Symptom**: AI references old, incorrect, or obsolete information

    **Causes:**

    * Memory not updated with new information
    * Conflicting facts stored
    * Manual memory editing needed

    **Solutions:**

    1. Navigate to Memory settings
    2. Review stored information
    3. Delete outdated entries
    4. Add corrected information
    5. Confirm changes are applied

    See [managing Memory](/memory/manage-reset-export) for detailed instructions.
  </Accordion>

  <Accordion title="Important details not remembered">
    **Symptom**: AI forgets critical facts you've shared multiple times

    **Causes:**

    * Information not structured for retention
    * Memory feature disabled
    * Details not deemed significant by AI
    * Storage limits reached

    **Solutions:**

    1. **Explicitly state importance**: "Remember this for future reference..."
    2. **Use structured format**: Present information as facts or lists
    3. **Verify memory storage**: Check Memory settings to confirm storage
    4. **Repeat in multiple conversations**: Reinforce important details
    5. **Use custom instructions**: Add to [custom instructions](/chat/custom-instructions) instead
  </Accordion>
</AccordionGroup>

## Understanding context windows

<Tabs>
  <Tab title="How context works">
    AI models process conversations within a **context window** - a limited amount of text they can reference at once.

    **What happens when limit is reached:**

    1. Oldest messages are dropped
    2. AI can't reference removed content
    3. May appear to "forget" earlier discussion

    **Typical context windows:**

    * GPT-4 Turbo: 128K tokens (\~96,000 words)
    * Claude 3.5 Sonnet: 200K tokens (\~150,000 words)
    * Gemini 1.5 Pro: 2M tokens (\~1.5M words)
    * GPT-4o: 128K tokens (\~96,000 words)

    <Info>
      One token ≈ 0.75 words on average. Conversation includes both your messages and AI responses.
    </Info>
  </Tab>

  <Tab title="Managing context">
    **Efficient context management:**

    1. **Start new chats for new topics**: Don't mix unrelated discussions
    2. **Use Memory for persistent facts**: Store information that spans conversations
    3. **Summarize long discussions**: Periodically create summaries
    4. **Remove unnecessary files**: Uploaded files consume context
    5. **Archive old conversations**: Keep active chats focused

    **When to start fresh:**

    * After 50+ message exchanges
    * When switching topics
    * If AI shows confusion
    * After completing a project phase
  </Tab>

  <Tab title="Token optimization">
    **Reduce token usage:**

    <CodeGroup>
      ```text Inefficient (high tokens) theme={null}
      Please help me understand the comprehensive details about how machine learning algorithms work, including neural networks, decision trees, random forests, and all the mathematical foundations behind each one, with detailed explanations of backpropagation, gradient descent, loss functions, and optimization techniques...
      ```

      ```text Efficient (low tokens) theme={null}
      Explain key machine learning concepts: neural networks, decision trees, and gradient descent. Focus on practical understanding.
      ```
    </CodeGroup>

    **Optimization tips:**

    * Be concise in follow-ups
    * Avoid repeating information
    * Use bullet points over paragraphs
    * Reference earlier messages instead of rewriting
  </Tab>
</Tabs>

## Fixing memory issues

<Steps>
  <Step title="Enable and verify Memory">
    1. Navigate to Settings → Memory
    2. Enable the Memory feature
    3. Confirm it's active (indicator should show enabled)
    4. Test by asking "What do you remember about me?"

    <Check>
      Memory is working when AI can recall facts from previous conversations.
    </Check>
  </Step>

  <Step title="Review stored memories">
    1. Open Memory management page
    2. Read through stored information
    3. Verify accuracy of remembered facts
    4. Look for conflicts or outdated data

    <Frame>
      <img src="https://mintlify.s3.us-west-1.amazonaws.com/reedai-07fa30f1/images/memory-management.png" alt="Memory management interface" />
    </Frame>
  </Step>

  <Step title="Clean up memory">
    Remove or update incorrect information:

    1. Click edit on any memory entry
    2. Correct the information or delete entirely
    3. Add new facts if needed
    4. Export backup before major changes

    See [memory management guide](/memory/manage-reset-export) for details.

    <Warning>
      Resetting memory deletes all stored information permanently. Export first if you want to preserve anything.
    </Warning>
  </Step>

  <Step title="Optimize conversation structure">
    For complex projects, structure conversations effectively:

    **Project chat organization:**

    ```
    Main project chat: High-level decisions and planning
    ├── Implementation chat 1: Feature A development  
    ├── Implementation chat 2: Feature B development
    └── Bug fixes chat: Issues and resolutions
    ```

    Use [conversation branches](/chat/continue-and-branches) to explore alternatives without cluttering main discussion.
  </Step>
</Steps>

## Memory feature best practices

### What to store in Memory

<AccordionGroup>
  <Accordion title="✅ Good candidates for Memory">
    * Personal preferences and work style
    * Project requirements and specifications
    * Technical environment details (languages, frameworks)
    * Business context and domain knowledge
    * Recurring tasks and patterns
    * Team structure and responsibilities
    * Important deadlines and milestones

    **Example:**

    ```text theme={null}
    "I work primarily in React and TypeScript"
    "My projects use PostgreSQL database"
    "I prefer functional programming patterns"
    "My team follows Agile methodology"
    ```
  </Accordion>

  <Accordion title="❌ Not suitable for Memory">
    * Sensitive personal information
    * API keys, passwords, or secrets
    * Constantly changing data
    * Temporary project details
    * Conversation-specific context
    * Very detailed technical specifications

    <Warning>
      Never store passwords, API keys, tokens, or other sensitive credentials in Memory. Use secure credential management instead.
    </Warning>
  </Accordion>
</AccordionGroup>

### Structuring information for retention

**Effective memory statements:**

* Clear and concise
* Factual, not conversational
* Specific and actionable
* Updated when circumstances change

<CodeGroup>
  ```text ❌ Poor structure theme={null}
  We talked about how I might want to maybe use Python sometimes for data stuff when I'm working on projects that involve data analysis or something similar to that.
  ```

  ```text ✅ Good structure theme={null}
  I use Python for data analysis projects.
  Preferred libraries: pandas, numpy, matplotlib.
  ```
</CodeGroup>

## Advanced troubleshooting

<Accordion title="Memory not persisting across sessions">
  **Possible causes:**

  * Browser cookies disabled
  * Using incognito/private mode
  * Different device/browser
  * Account sync issues

  **Solutions:**

  1. Enable cookies in browser settings
  2. Use regular browsing mode
  3. Check account sync status
  4. Try logging out and back in
  5. Contact support if issue persists
</Accordion>

<Accordion title="Conflicting memories causing confusion">
  When Memory contains contradictory information:

  1. **Identify conflicts**: Review all stored memories
  2. **Prioritize current info**: Delete outdated facts
  3. **Be explicit in prompts**: Override Memory when needed
  4. **Reset if necessary**: Start fresh if too complex

  **Overriding Memory in prompts:**

  ```text theme={null}
  Despite what you may remember about my language preference, 
  for this project I'm using Go instead of Python.
  ```
</Accordion>

<Accordion title="Privacy concerns with Memory">
  Memory is stored securely and never shared:

  * Encrypted at rest and in transit
  * Only accessible to your account
  * Not used for model training
  * Can be exported or deleted anytime

  See [data and compliance](/privacy/data-and-compliance) for details.

  <Info>
    You can export or delete all Memory data from your account settings at any time.
  </Info>
</Accordion>

## Related pages

<CardGroup cols={2}>
  <Card title="How Memory works" icon="brain" href="/memory/how-memory-works">
    Understand Memory feature in depth
  </Card>

  <Card title="Manage Memory" icon="sliders" href="/memory/manage-reset-export">
    Edit, export, and reset Memory
  </Card>

  <Card title="Token limits" icon="gauge" href="/prompts/tokens-and-limits">
    Understanding context windows
  </Card>

  <Card title="Custom instructions" icon="wand-magic-sparkles" href="/chat/custom-instructions">
    Set persistent conversation guidelines
  </Card>
</CardGroup>