Hussein Khidr

I'm a founder in the depths of pivot hell. But I have a strong feeling that the important space to build in is memory. For me, the thing that can increase the value we generate from AI 10x from where the models currently are, even without further improvement, is memory and context. Being able to transfer context from one interaction to the other, preferences for how responses should be structured, permission management, sensitive information filtering and redacting, etc. Of course, all these have to be done securely and end-to-end.
I've thought a lot about how to find this wedge. A lot of things I've found just sound cool but probably aren't providing much value to the user. I'm optimizing for a high-value, high-frequency use case. This is the best-case scenario. Especially for a consumer-based solution. I would list some of the solutions that I think can stem from having an end-to-end encrypted memory of all AI interactions (limited to conversations with AI models for now) and a system that does a basic filtering of sensitive information and PIL:

• Smart search: being able to semantically search through old conversation history to find working solutions with pure conversation.

• Prompt injection: Being able to use old verify prompts with a click to ask better questions.

• Conversation summaries: Just as it implies, get summaries of conversations, perhaps of just one chat or multiple. And maybe even across multiple AI tools.

• Project workspaces: Having an organization share a connected memory. With each team member being able to use the other members' context on what they're working on to make sure they're all aligned on the goal and keep a coherent implementation of the tasks.

• Chat with memory: Using a lightweight LLM to converse with your chat history across all AI use and different conversations and sessions.

I'm a founder in the depths of pivot hell. But I have a strong feeling that the important space to build in is memory. For me, the thing that can increase the value we generate from AI 10x from where the models currently are, even without further improvement, is memory and context. Being able to transfer context from one interaction to the other, preferences for how responses should be structured, permission management, sensitive information filtering and redacting, etc. Of course, all these have to be done securely and end-to-end.

It's always easy to see the long-term play or vision. The easiest to recognize here is that as more and more AI tools get built, it would be great to have a way for the user to log in once and get memory and context from all other AI interactions transferred at once without having to take time to build. This would probably be done by providing APIs or SDKs to developers to integrate with, and then we can act like a kind of AI app store and recommend AI apps that are fully integrated with context whenever the user needs it.

I'm a founder in the depths of pivot hell. But I have a strong feeling that the important space to build in is memory. For me, the thing that can increase the value we generate from AI 10x from where the models currently are, even without further improvement, is memory and context. Being able to transfer context from one interaction to the other, preferences for how responses should be structured, permission management, sensitive information filtering and redacting, etc. Of course, all these have to be done securely and end-to-end.

It's always easy to see the long-term play or vision. The easiest to recognize here is that as more and more AI tools get built, it would be great to have a way for the user to log in once and get memory and context from all other AI interactions transferred at once without having to take time to build. This would probably be done by providing APIs or SDKs to developers to integrate with, and then we can act like a kind of AI app store and recommend AI apps that are fully integrated with context whenever the user needs it.