Advanced control features
Last updated
Was this helpful?
Last updated
Was this helpful?
Memex is always willing and ready to do work for you, but there might be use-cases or specific workflows where you might preffer to have more fine-grained control over the workflow.
There are three core advanced control features: Code Execution Approval, Turns Control, and Thinking.
These features can be managed from the settings panel (wrench icon) under the conversation input. They add a whole new layer of control to Memex, helping you ensure it doesn't do any unwanted work.
This features gives you the ability to have Memex ask for approval before running each code block. When turned on, you'll be able to review each code block to make sure it aligns with your expectations, and select if you'd like Memex to run it or skip it so you can provide more guidance.
After you choose to Run or Skip a specific code block, Memex will add a small visible note to let you know which action you selected if you need to review your work. This note also helps Memex know what action it took with that block, and inform its next steps accordingly.
With this feature you can set a maximum number of turns Memex can go through autonomously between user commands before stopping. This can be helpful when you want tighter control on specific parts of your projects, or just generally prefer to be more in the loop of the workflow.
The number of turns resets whenever there is a user message or a new conversation is started.
The Long Context Mode addresses a fundamental challenge with AI coding assistants: maintaining project context across long conversations.
Long Context Mode is powered by Gemini 2.5 Pro, offering a 1M context window with plans to expand to 2M tokens. We've found Gemini particularly strong for system-level tasks, debugging, and quantitative work.
When you hit context limits with Claude, you can seamlessly switch to Gemini and continue working.
Our internal testing shows Gemini 2.5 Pro excels at:
Complex debugging scenarios
System-level programming tasks
Quantitative analysis and computational logic
Long-running conversations that require maintaining context
If you find yourself stuck on a stubborn bug with another model, switching to Gemini 2.5 Pro significantly increases the chances of a swift resolution. While Claude excels at file editing, Gemini's strength lies in generating sophisticated implementations, even if it sometimes requires a quick syntax check.
If you are experiencing degradation in file editing or in general, you can turn off Long Context mode to get back to standard performance.
Click on the 🔧 icon
Toggle the Long context mode OFF
Thinking Mode enables Memex to engage in explicit reasoning when addressing your requests. Rather than simply generating direct answers, when Thinking Mode is activated, Memex deliberately works through problems step-by-step, breaking down complex tasks into a series of logical intermediate reasoning steps - similar to how humans approach problem-solving.
When enabled, not only does Memex engage in reasoning, it also makes the process visible, allowing you to see how it interprets information, weighs different considerations, and arrives at conclusions. This transparency provides valuable insight into Memex's decision-making process and helps build understanding of how it approaches your requests.
You can turn Thinking Mode on from the wrench settings menu. When on, you can select from four Thinking Levels: Level 1 provides basic reasoning for simpler tasks, while Level 4 offers detailed, comprehensive reasoning for complex problems requiring thorough analysis.
Although Thinking Mode might increase credit usage in an individual message, we've found it leads to more efficient results overall. By prompting Memex to follow a structured reasoning path, responses become more focused, accurate, and grounded in your specific task. The step-by-step approach reduces tangential responses and helps Memex avoid jumping to premature conclusions. It also creates natural pause points, preventing Memex from continuing unnecessarily when your request has been adequately addressed.
