AI Memory

AI Memory allows the AI to save important facts about you during conversations and recall them automatically in future interactions. This creates continuity across threads, enabling the AI to personalize its responses based on accumulated knowledge about your preferences, context, and goals.

Overview

When you interact with the AI, it can recognize information worth remembering -- your name, your role, your preferred communication style, project context, or key decisions. The AI saves these facts using a built-in memory tool, and they are automatically included in every subsequent conversation. You do not need to re-state your preferences or background in each new thread.

Memory is a system tool, meaning it is available by default in all threads without requiring explicit configuration. The AI decides when to save a memory based on the conversational context.

How It Works

Saving Memories

The AI uses the memory tool to save facts during conversations. When you share information that the AI determines is worth remembering, it calls the memory tool with the relevant fact. This appears as a tool call in the conversation, making the process fully transparent.

What triggers a memory save:

• You mention personal preferences ("I prefer concise responses")
• You share professional context ("I'm a backend engineer working on a NestJS project")
• You state goals or objectives ("We're preparing for a product launch in Q2")
• You provide key facts ("Our company uses PostgreSQL and Redis")
• You express communication preferences ("Always respond in French")

When the AI saves a memory, you see a tool call in the response indicating what was remembered. For example:

Tool: memory
Input: { "content": "User prefers TypeScript over JavaScript for all new code" }
Result: Remembered: "User prefers TypeScript over JavaScript for all new code"

Automatic Recall

Saved memories are automatically injected into the system prompt of every conversation. You do not need to request recall or reference memories explicitly. The AI receives all your saved memories as context before processing your message, allowing it to tailor its responses accordingly.

The memories are formatted as a structured list within the system prompt:

Facts remembered about this user:
- User prefers TypeScript over JavaScript for all new code
- User's name is Marie, she is a senior developer at Acme Corp
- User's team follows the Angular style guide for naming conventions
- User prefers dark mode examples in documentation

Memory Scope

Memories are per-user. Each user has their own set of memories that are recalled only in their conversations. Memories are not shared with team members or visible to other users.

Key scoping rules:

• Memories persist across all threads for the same user
• Memories are not thread-specific -- they apply globally
• Memories are not team-specific -- they follow the user across teams
• Each memory is a standalone fact (not hierarchical)

What Gets Saved

The AI is instructed to save the following categories of information:

Category	Examples
Personal preferences	Language preference, response style, formatting preferences
Professional context	Role, company, team, tech stack, industry
Goals and objectives	Current projects, deadlines, strategic priorities
Key facts	Names, tools used, standards followed, past decisions
Communication style	Preferred tone, level of detail, language

The AI exercises judgment about what to save. Not every piece of information warrants a memory. Trivial or one-time details are typically not saved unless they indicate a broader preference.

Privacy and Control

Transparency

Every memory save is visible in the conversation as a tool call. The AI does not save information silently. You can see exactly what was remembered and when.

Reviewing Memories

Your saved memories can be reviewed through your account settings. The memory management interface displays all stored facts with their creation dates, allowing you to audit what the AI knows about you.

Deleting Memories

Individual memories can be deleted if they are no longer accurate or desired. You can also clear all memories at once to start fresh. Deleted memories are permanently removed and will not be included in future conversations.

To manage memories:

1. Navigate to your account settings
2. Open the Memory section
3. Review the list of saved facts
4. Delete individual memories or clear all

You can also instruct the AI conversationally to forget specific information. For example, saying "Forget that I prefer Python" will prompt the AI to acknowledge the change, though for permanent removal, use the settings interface.

Use Cases

Personalized Assistance

A marketing manager who regularly asks for content suggestions can have the AI remember brand voice guidelines, target audience demographics, and preferred content formats. Every new content request automatically incorporates these preferences without restating them.

Development Continuity

A developer working across multiple threads on the same codebase can have the AI remember the project's technology stack, architectural decisions, and coding conventions. When starting a new thread for a different module, the AI already knows the project context.

Cross-Thread Knowledge

A consultant managing multiple client engagements can have the AI remember key facts about each client. When switching between threads for different clients, the shared memories provide baseline context that reduces the need for repetitive briefing.

Team Onboarding

New team members can interact with the AI and gradually build up a memory profile that reflects their role, responsibilities, and preferences. Over time, the AI becomes more effective as it accumulates relevant context.

Technical Details

Storage

Memories are stored in the user_memories table with the following structure:

Field	Description
ID	Unique identifier
User ID	The user who owns the memory
Content	The fact text
Created At	When the memory was saved
Updated At	Last modification timestamp

Injection

Memories are assembled into a formatted block and injected into the system prompt before each AI request. The injection happens during the history building phase, ensuring the AI has access to all saved memories before processing the user's message.

Capacity

There is no hard limit on the number of memories, but each memory consumes tokens from the context window. Users with hundreds of memories may notice increased token usage in the system prompt. In practice, the AI tends to save concise facts, so memory token consumption remains modest (typically 100-500 tokens for an active user).

Best Practices

1. Let the AI decide -- The AI is designed to recognize memorable information organically. You do not need to explicitly ask it to remember things, though you can if a fact is particularly important.
2. Be explicit about preferences -- If you have strong preferences about response format, language, or style, state them clearly. The AI is more likely to save explicitly stated preferences.
3. Review periodically -- Check your saved memories occasionally to remove outdated facts. Role changes, project completions, or evolving preferences should be reflected in your memory profile.
4. Correct inaccuracies -- If the AI saves something incorrect, tell it directly. You can also delete the memory from settings and re-state the correct information.
5. Use for stable facts -- Memories work best for information that is stable over time (preferences, roles, standards). For information that changes frequently, context blocks attached to specific threads are more appropriate.
6. Monitor token impact -- If you notice your system prompt token usage increasing, review your memories. Consider deleting obsolete facts to keep the memory footprint lean.

Related Documentation

• System Prompts - How system prompts work with injected memories
• Context Blocks - Thread-specific context vs. global memories
• Tools Overview - Understanding system tools and tool calls
• Threads - Thread management and conversation continuity