Skip to main content
Memory is the core capability that enables Expert personalization. Unlike typical chatbots that “forget everything once you close the window,” Profy’s Experts can remember your preferences, decisions, and context across sessions. Every conversation accumulates knowledge, making your Expert understand you better over time.

​
Three Layers of Memory

Profy’s memory system consists of three layers, each serving a different role:

Memory List

User-visible memory entries. You can view, edit, disable, or delete them — this is the “source of truth” for memory

Core Memory

Key knowledge distilled by AI from the memory list. Organized as entities and facts, injected into every conversation

Memory Graph

A visual projection of core memory. Displays relationships between entities and timelines in graph form

​
Where Memory Comes From

​
Automatic Extraction

This is the primary source of memory. After each conversation, the system automatically analyzes the dialogue and extracts valuable information:
  • Preferences you express: “I prefer concise code style” “Please reply in English”
  • Decisions you make: “We’ll go with Option A” “Use React instead of Vue”
  • Important facts: “My project uses Python 3.11” “The deadline is next Friday”
Not all conversation content gets recorded. The system has quality gates — only information that is specific enough, durable, and non-obvious will be extracted.

​
Manual Memory

You can also explicitly tell the Expert: “Remember, my team has a standup every Wednesday.” The Expert will save this information using its memory tool.

​
Creator Presets

Expert creators can preset foundational memories during the distillation process, ensuring the Expert has basic domain knowledge from the very first conversation with any user.

​
How Memory Works

1

Conversation Accumulation

During your conversations with the Expert, the system records all interactions as raw material.
2

Automatic Extraction

After a conversation ends, AI analyzes the content and extracts memory entries that meet quality standards (specific enough, durable, non-obvious).
3

Knowledge Distillation

The system periodically performs “consolidation” on accumulated memory entries, distilling related information into more structured core memory: extracting entities, generalizing facts, and establishing relationships.
4

Injection into Conversations

In the next conversation, relevant memories are retrieved and injected into the Expert’s context. This is how the Expert “remembers” what you said before.

​
Managing Your Memory

You have full control over what the Expert remembers.

​
Viewing Memory

In the “Memory” tab on the Expert page, you can see three sub-views:

​
Editing Memory

You can modify the content of any memory entry. If the Expert remembered something incorrectly, simply edit it to correct it.

​
Disabling Memory

If a memory entry isn’t needed right now but might be useful later, you can disable it. Disabled memories won’t be injected into the conversation context, but can be re-enabled at any time.

​
Deleting Memory

If you want the Expert to completely forget something, you can delete that memory entry. Deletion is permanent and irreversible.
Deletion and disabling are different operations. Deletion is permanent; disabling is reversible. Please confirm your intent before choosing.

​
Memory Source Transparency

During conversations, when the Expert uses memory to answer your questions, the memory sources are annotated below the message. You can see which memories the Expert referenced and what type they are (preference, constraint, decision, etc.). This transparency lets you judge whether the Expert’s response is based on correct context. If the referenced memory is wrong, you can correct it directly in the memory list.

​
Core Memory and Knowledge Graph

Core memory is structured knowledge distilled by AI from your memory entries, composed of entities and facts:
Entities represent key concepts in memory. They can be:
  • People: colleagues, clients, partners you mention
  • Projects: work projects you’re involved in
  • Tools: software and tech stacks you frequently use
  • Organizations: companies or teams you belong to
  • Concepts: important concepts in your domain of expertise
  • And more types…
Facts are specific statements about entities, for example:
  • “User prefers TypeScript”
  • “Project X’s deadline is end of August”
  • “Team uses Feishu for daily communication”
Each fact has a temporal dimension — when it became valid and whether it’s still valid.
Relations describe connections between entities:
  • “User → uses → TypeScript”
  • “Project X → belongs to → Company Y”
  • “User → prefers → concise style”
In the memory graph, these relations are presented visually.

​
Memory Evolution (Dream)

The system periodically performs “consolidation” on the Expert’s memory — internally called Dream. The consolidation process includes:
  1. Syncing newly added memory entries as evidence
  2. Distilling new entities and facts from evidence
  3. Handling contradictions — when new facts conflict with old ones, the old fact’s validity period is closed and the new fact takes over
  4. Updating relationships between entities
The consolidation process is fully automated; you don’t need to trigger it manually.
Memory consolidation does not modify or delete your original memory entries. It only distills more structured knowledge from them to improve conversation quality.

​
Temporary Chat and Memory

If you enable “Temporary Chat” mode:
  • The Expert will not read any existing memory
  • Conversation content will not be extracted as new memory
  • This conversation is completely transparent to the memory system
This is suitable for handling sensitive information or scenarios where you don’t want the Expert to accumulate context.

​
Next Steps

Skills

Learn how skills give Experts domain-specific capabilities

Expert

Review the complete composition and lifecycle of an Expert