Agent Settings
Customize how Memex works to match your preferences and workflow needs.
Overview
Agent Settings give you fine-grained control over Memex's behavior, allowing you to tailor its approach to match your working style, project requirements, and comfort level with automation.
Access Agent Settings by clicking the wrench (🔧) icon in the interface. These settings affect how Memex operates during your current conversation and can be adjusted at any time.
Custom Instructions
Custom Instructions are your way to set consistent guidelines that Memex follows across all conversations. Think of them as extensions to Memex's system prompt that reflect your personal preferences and requirements.
What Custom Instructions Do
Custom Instructions ensure Memex behaves consistently according to your preferences without you having to repeat the same guidance in every conversation.
Examples of effective Custom Instructions:
For Visual Content:
For Audio Processing:
For Development Workflow:
For Communication Style:
For Technical Preferences:
Setting Custom Instructions
Click the wrench (🔧) icon to open settings
Navigate to the "Custom Instructions" section
Enter your guidelines in plain language
Click "Save" to apply across all future conversations
[Insert custom instructions setup walkthrough here]
Best Practices for Custom Instructions
Be Specific but Not Overly Restrictive
Focus on Consistent Preferences
Consider Your Skill Level
Model Selector
Choose which AI model powers your Memex conversations, balancing speed, capability, and cost.
Available Models
Memex offers three model tiers, each optimized for different use cases:
Note: We currently use the Claude family model to power all tiers. This might change over time—we’ll keep upgrading them so each option always uses the best available fit for that tier.
Fast (Claude Haiku)
Best for: Quick tasks, simple questions, rapid iteration
Strengths: Fastest response times, lowest credit usage
Ideal use cases:
Simple code explanations
Quick file edits
Routine tasks you've done before
Brainstorming and ideation
Balanced (Claude Sonnet)
Best for: Most everyday development work
Strengths: Excellent balance of speed and intelligence
Ideal use cases:
Building features from scratch
Debugging complex issues
Code refactoring
Multi-file changes
Genius (Claude Opus)
Best for: Complex problems requiring deep reasoning
Strengths: Most capable model with superior problem-solving
Ideal use cases:
Architecture decisions
Complex algorithmic challenges
Novel problems without clear solutions
Tasks requiring extensive context understanding
Credit Usage
Each model tier has different credit costs per turn:
Fast (Haiku)
$
Balanced (Sonnet)
$$
Genius (Opus)
$$$
The cost indicators appear in the model selector dropdown, helping you make informed choices about credit usage.
How to Switch Models
Look for the model selector chip in the message input toolbar (shows a CPU icon with the current model name)
Click the chip to open the dropdown menu
Select your preferred model from the list
The selected model will be used for your next message and subsequent turns
You can switch models at any time during a conversation. This is useful when you want to:
Start with Fast mode for exploration, then switch to Balanced for implementation
Escalate to Genius mode when you hit a particularly challenging problem
Switch back to Fast mode for quick follow-up questions
Tips for Model Selection
Start Fast, Escalate When Needed
Begin conversations with the Fast model for initial exploration. If the task proves more complex than expected, switch to Balanced or Genius mid-conversation.
Match Model to Task Complexity
Simple tasks: Use Fast - no need for extra capability on straightforward work
Standard development: Use Balanced - it handles most coding tasks excellently
Hard problems: Use Genius - when you need the best reasoning available
Consider Your Budget
If you're working on a project with many interactions, using Fast for routine tasks helps preserve credits for when you really need the more capable models.
Combine with Thinking Mode
For complex problems, pairing Genius mode with Thinking Mode (Level 3-4) can produce exceptional results, though at higher credit cost.
Code Execution Control
Control when and how Memex executes code on your system.
Code Execution Approval
This feature requires Memex to ask for permission before running each code block, giving you complete control over what gets executed on your machine.
When enabled:
Memex shows you each code block before execution
You can choose to "Run" or "Skip" each block
Memex waits for your approval before proceeding
Visual indicators show which blocks were run or skipped
When to use Code Execution Approval:
Working with sensitive or critical systems
Learning how specific implementations work
When you want to review each technical step
During debugging to control execution flow
Tips for effective use:
Review code blocks for understanding, not just errors
Use "Skip" when you want to provide different guidance
Code approval helps you learn implementation patterns
Creates natural pause points for questions or modifications
Max Turns Control
Limit how many autonomous actions Memex can take before requiring your input.
How Max Turns Works
By default, Memex continues working autonomously until it completes your request. Max Turns lets you set a limit on how many iterations it can go through before stopping to check with you.
Turn counting:
Each code execution counts as one turn
Each file modification counts as one turn
Web searches and API calls count as turns
The counter resets when you send a new message
When to Use Max Turns
For Tight Control: Set to 1-3 turns when you want to review every action:
Memex will execute at most 2 actions before stopping for your guidance.
For Moderate Oversight: Set to 5-10 turns for balanced automation with oversight:
Good for complex tasks where you want regular check-ins.
For Full Automation: Set to unlimited (default) when you trust Memex to complete the entire task:
Effective Max Turns Strategies
Learning Mode: Use low turn counts when learning how Memex approaches problems:
You'll see each step of the implementation process.
Review Mode: Use moderate turn counts for complex projects:
Regular checkpoints without micromanagement.
Production Mode: Use unlimited turns for well-defined tasks:
Let Memex work efficiently without interruption.
Long Context Mode
Extended context capabilities powered by advanced AI models.
What Long Context Mode Provides
Long Context Mode addresses a fundamental challenge: maintaining project understanding across extended conversations. It's powered by Claude's Sonnet 4 1M token context window, with plans to expand further.
Capabilities:
Handle much longer conversations without losing context
Maintain detailed project understanding across multiple sessions
Better performance on complex, multi-part projects
Reduced need for conversation summaries
When to Enable Long Context
Large Projects: Enable for projects with extensive codebases or complex requirements.
Extended Development Sessions: When you plan to work for extended periods without breaks.
Complex Context Requirements: When project success depends on maintaining detailed context.
Note: Usage beyond the regular context window of 200k might result in higher credit usage per turn, in alignment with Claude's Sonnet 4 Long Context pricing structure
Thinking Mode
Enable explicit reasoning for better problem-solving and transparency.
How Thinking Mode Works
Instead of jumping directly to solutions, Thinking Mode enables Memex to work through problems step-by-step, showing you its reasoning process.
What you see with Thinking Mode:
Step-by-step problem analysis
Consideration of different approaches
Reasoning behind technical decisions
Transparent thought process
Thinking Levels
Choose the depth of reasoning that matches your task complexity:
Level 1: Light
Simple problem breakdown
Quick decision explanations
Suitable for straightforward tasks
Level 2: Balanced
More thorough problem analysis
Multiple approach consideration
Good for typical development tasks
Level 3: Deep
Comprehensive problem exploration
Detailed tradeoff analysis
Ideal for complex technical decisions
Level 4: Max
Extensive reasoning and exploration
Multiple solution paths considered
Best for challenging problems requiring thorough analysis
[Insert thinking mode walkthrough here]
Benefits of Thinking Mode
Better Outcomes: Structured reasoning leads to more thoughtful, well-considered solutions.
Learning Opportunity: See how expert-level thinking approaches technical problems.
Quality Assurance: Transparent reasoning helps you evaluate the approach before implementation.
Debugging Aid: When things go wrong, reasoning traces help identify where issues occurred.
Credit Efficiency
Despite using more tokens per message, Thinking Mode often reduces overall credit consumption by:
Producing more accurate solutions on first attempt
Reducing need for iteration and debugging
Creating natural stopping points
Avoiding premature or incorrect implementations
Task Management (Beta)
Enable structured task tracking for plan implementation.
What Task Management Does
Task Management is an experimental feature that enables Memex to create and maintain structured task lists when implementing project plans. When enabled, Memex can automatically break down plans into discrete, trackable tasks and systematically work through them. This feature is currently in beta as we refine its behavior and integration with different project workflows.
Important: Due to the experimental nature of this feature, it's recommended to explicitly ask Memex to "use task management" in your prompt even after enabling it. This helps ensure more reliable activation as we continue to improve the feature's behavior and integration.
Note: Behavior may change as we gather feedback and improve the system.
Agent Settings Best Practices
For Beginners
Custom Instructions: "Explain technical concepts clearly and provide step-by-step guidance"
Code Execution: Approval enabled for learning
Max Turns: 3-5 for oversight
Thinking Mode: Level 2 for educational benefit
For Experienced Users
Custom Instructions: Focus on your specific tech stack preferences
Code Execution: Auto-execute for speed
Max Turns: 10+ or unlimited
Thinking Mode: Level 3-4 for complex problems only
For Team Projects
Custom Instructions: Include team coding standards and practices
Code Execution: Approval for shared environments
Max Turns: Moderate (5-7) for reviewable progress
Long Context: Enabled for project continuity
Project-Specific Adjustments
Experimental Projects:
Higher turn limits for exploration
Thinking Mode for learning new approaches
Less restrictive execution controls
Production Projects:
Code approval for critical systems
Documentation requirements in Custom Instructions
Conservative turn limits for safety
Agent Settings give you the flexibility to make Memex work exactly how you want it to. Experiment with different configurations to find what works best for your style and projects.
Last updated
Was this helpful?