Working with long conversation

Model context window limit

LLMs work using context windows, which represent how much input an LLM can process at once. Claude 3.7, the LLM Memex uses, has a context window of 200k tokens.

You can see how much of context window a conversation is using in the usage indicator at the bottom right of the app:

Memex helps you deal with potentially running out of context via its Summary for New Conversation feature.

Creating a Summary for a New Conversations

By creating a summary of your existing conversation, with special focus on its technical details, you are able to create a new conversation that is able to continue the work from a previous one.

You can prompt Memex to generate this type of summary for you by on the Context Management icon on the bottom right, and clicking "Generate summary to start new conversations".

After Memex generates the summary, you can select "Start a new conversation", which will set the summary as a prompt in a new conversation in the same project (important so it can utilize the same resources and environment).

As you start getting close to the context limit, Memex will also prompt you use this feature to help you continue your work.

Starting a new conversation in the same project

You can start a conversation in a specific project by starting it from the Project page, which can be reached by clicking on the project name in the path at the top of the app, or by clicking on the project from the Home Screen.

Context Usage Management

Context Usage is an experimental feature that allows you to control how much conversation history Memex includes in each interaction, balancing context awareness with conversation length. You can select from four options:

• Small Window

• Medium Window

• Large Window

• Full History

When using a window size, Memex combines:

• The complete raw content from your most recent messages (within the selected window)

• A dynamic summary of the rest of the conversation (outside the selected window)

This approach helps Memex focus on the latest turns by replacing older raw conversation with condensed summaries while maintaining contextual awareness, enabling you to have longer conversations, as each turn has less raw input. Choose your window size based on your project needs:

• Larger windows provide more detailed context but use a higher percentage of the model context limit

• Smaller windows allow for longer conversations and are suitable when older context is less relevant to current tasks

Tip: If you are about to run out of context limit, set your window to Large. This might be counter-intuitive, as this window size mantains a larger % of the raw messages, but the rest of the conversations that is being summarized also has to fit in the model to be summarized. A Small window keeps a lower % of the raw messages, which can lead to the context that needs to be summarized not fitting in the model context.

Best practices

When working on big project, put the following to your Custom Instructions or append to your prompt

Create a step by step plan before proceeding Document your progress in README.md Commit code after each successful step Confirm with me after each milestone

Then after each significant milestone in the current conversation, start a new conversation in the same project and tell Memex

Pick up README and start/continue working on feature X

Following these best practices will help with 2 things:

• Picking up context for new convo, trying different path, etc

• Shorter conversations also translate to better credits efficiency - the intuition is, the longer the convo is, the more expensive each turn of interaction with the agent are.

In order for Memex to pick up on the project context, if you haven't asked it to write docs/readme in the previous conversation, you can try something like this prompt

I have an existing project X which does Y, in the current project directory. Take a look around to understand the context, tech stack, current implementation. Then start working on next feature Z or continue fixing issue Z"