> ## Documentation Index
> Fetch the complete documentation index at: https://docs.profy.cn/llms.txt
> Use this file to discover all available pages before exploring further.

# Expert Memory

> View and manage an Expert's memory list, core memory, and memory graph

Each Expert has its own independent memory system for accumulating specialized knowledge learned through interactions with you. Unlike [personal memory](/en/documentation/memory/personal-memory) (which focuses on "who you are"), Expert memory focuses on "what the Expert has learned while serving you"—your project details, approaches discussed, decisions made, and pitfalls encountered.

## Memory Panel

In an Expert's detail page, you'll find the memory panel. The panel contains three views that present the Expert's accumulated knowledge from different perspectives.

<Tabs>
  <Tab title="Memory List">
    All stored memory entries displayed as a list. Each memory is a specific fact or piece of knowledge derived from your conversations with this Expert.

    **Supported operations**:

    * **Browse**: View each memory's content and source
    * **Enable / Disable**: A reversible operation to temporarily turn off a memory—disabled memories won't be referenced by the Expert
    * **Delete**: Permanently remove a memory (irreversible)

    This is the most straightforward memory management view, ideal for quickly reviewing and adjusting specific entries.
  </Tab>

  <Tab title="Core Memory">
    Key knowledge distilled by AI from the memory list, organized as entities and relationships.

    Core memory is a "highly condensed" representation of the Expert's knowledge about you. It automatically consolidates numerous fragmented memory entries into structured knowledge:

    * **Entities**: Key objects identified (people, projects, technologies, preferences, etc.)
    * **Facts**: Specific statements about entities (e.g., "User's project uses PostgreSQL as the primary database")
    * **Relationships**: Connections between entities (e.g., "Project A uses Framework B")

    Core memory is the knowledge source the Expert actually references during conversations—it extracts the most valuable insights from raw memories and injects them into the conversation context in a structured way.
  </Tab>

  <Tab title="Memory Graph">
    A visual graph showing the network of relationships between pieces of knowledge.

    The memory graph renders the entities and relationships from core memory as a network diagram: nodes represent entities, and edges represent their relationships. You can intuitively see the full picture of the Expert's knowledge about you—which concepts are central nodes and how they connect.

    This view is primarily for **understanding and review**, helping you assess whether the Expert's knowledge about you is accurate and complete.
  </Tab>
</Tabs>

## How Memory Is Built

Expert memory accumulates through two mechanisms:

### Conversation Extraction

After each conversation, the system analyzes the exchange in the background and extracts valuable information into the memory list. Extraction follows strict quality standards—only specific, durable, and non-obvious facts are saved.

<Tip>
  You can also proactively tell the Expert what to remember during a conversation. For example, "Remember: this project's API version must remain v2-compatible."
</Tip>

### Memory Consolidation (Dream)

Beyond per-conversation extraction, the system also periodically "consolidates" existing memories. This process is similar to how human memory consolidation works:

<Steps>
  <Step title="Collecting Evidence">
    Raw information fragments extracted from conversations are archived as "evidence," with each piece of evidence pointing to its specific conversation source.
  </Step>

  <Step title="Distilling Core Knowledge">
    AI analyzes all evidence and distills structured core facts—identifying key entities, their relationships, and consistency or contradictions with your existing knowledge.
  </Step>

  <Step title="Handling Contradictions">
    When new information contradicts existing memory (e.g., you changed your technology stack), the old fact is marked as expired and the new fact takes over. Historical records are never deleted, preserving the complete evolution trail.
  </Step>

  <Step title="Updating the Graph">
    Consolidated knowledge is reflected in the core memory and memory graph views, allowing you to see how the Expert's understanding of you evolves over time.
  </Step>
</Steps>

<Note>
  Memory consolidation runs automatically in the background and does not affect your experience. The process never modifies the original entries in your managed memory list.
</Note>

## Three-Layer Architecture

The Expert's memory system consists of three layers, progressively refined from raw to distilled:

| Layer                   | Description                                                                                      | Visible to you?                                               |
| ----------------------- | ------------------------------------------------------------------------------------------------ | ------------------------------------------------------------- |
| **Memory List**         | Fact entries extracted directly from conversations—the raw form of memory                        | ✅ Fully visible and manageable                                |
| **Evidence**            | Immutable conversation-source snapshots for each memory, ensuring traceability                   | Indirectly visible (via source annotations on memory entries) |
| **Core Memory / Graph** | A structured knowledge network distilled by AI—the knowledge base the Expert actually references | ✅ Viewable (Core Memory tab / Graph tab)                      |

<Warning>
  Core memory and the graph are automatically generated—you cannot directly edit them. If you find an inaccurate fact in core memory, correct the corresponding original entry in the "Memory List" (disable or delete it), and the system will update core memory during the next consolidation.
</Warning>

## What Memory Does

The Expert loads its memory context at the start of each conversation. This memory enables the Expert to:

* **Continue previous work**: No need to re-introduce your project background every conversation
* **Maintain consistency**: Remember technical approaches and design decisions you've confirmed
* **Avoid repeating mistakes**: Remember pitfalls encountered and their solutions
* **Understand your preferences**: Know your code style, communication habits, and quality standards

## Management Tips

<AccordionGroup>
  <Accordion title="Periodically Review the Memory List">
    It's a good idea to browse the Expert's memory list from time to time to ensure the stored information is still accurate. As projects evolve, earlier technical decisions may have changed.
  </Accordion>

  <Accordion title="Proactively Correct">
    If an Expert references outdated memory during a conversation (e.g., mentioning an approach you've already abandoned), simply tell it "That's changed—here's the current situation…" The Expert will update its memory.
  </Accordion>

  <Accordion title="Leverage the Core Memory View">
    Core memory presents the Expert's understanding of you in a structured way. If you notice that a key entity (such as your main project) is missing or inaccurately described in core memory, proactively provide the relevant information in a conversation.
  </Accordion>

  <Accordion title="Use the Graph for Big-Picture Review">
    The memory graph is best suited for a "bird's-eye view" of the Expert's knowledge system. When you feel the Expert's understanding of a certain area is incomplete, use the graph to quickly locate knowledge gaps and then fill them in through targeted conversations.
  </Accordion>
</AccordionGroup>

## Creators and Memory

If you are an Expert's creator, you can also initialize the Expert's memory through:

* **Skill injection**: Write skill documents to inject domain knowledge into the Expert
* **Distillation conversations**: Inject core knowledge through in-depth dialogue during Expert creation

These creator-injected memories work alongside user-conversation-generated memories within the same system.