Introducing The Memory Feature

Journalling has stayed the same for a pretty long time. You take a pen and paper, and write down your thoughts. Now over time, we've add a couple features like tags, dates, and sharing your journal entries. But the main flow for journalling hasn't really seen any change.

This is where AI comes in. Now in many cases, AI starts to have a negative effect on things, eg. Imagine if you had an AI write a journal entry for you. Now that's not useful and it's actually bad since you're not getting the real benefits of journalling. Our philosophy is that AI should be a tool to support the user, not replace the user.

So, we created `Memory`. The idea is to have the AI learn just like human memory does so it can generate helpful insights, and information like an AI is meant to do.

The Idea

Let's say you've written your first journal entry and it's about your first day at work. You've met a couple key people, written down about your thoughts and feelings and had a couple interactions. But chances are, you'll probably never read this journal entry again because the fact is, people are busy with present and the future.

But you click the "Memorise" button, within a couple seconds, an AI has scanned through your journal entry, extracted key people, events and used it's knowledge about psychology to generate a profile of you.

Now, it's been a week, and you write a new journal entry. This time you discuss how you've adapted to your work, made a couple friends, etc. You click the "Memorize" button again, and the AI starts going through the memory process again. But this time, we pass your initial generation as context.

Over time, as you keep memorising journal entries, the AI grows more confident and understands you better. Just like human memory, events deemed no longer important are lost, while existing information is strengthened.

Now you need a way to interact with this "memory", so we have an AI Chat feature that you can talk to. Ask it questions or explore thoughts and the AI will fetch context from your memorised journal entries, your memory, your goals etc. Now you have a chatbot whose purpose is to elevate your understanding about yourself and promote your progress towards your goals.

Understanding the limitations

As great as AI is, it's not perfect. It's expensive to run, hallucinations can happen and the fact is, AI isn't a human. It's great for making sense of some data but an AI is going to understand human emotions, thoughts, experiences etc, at least not to the extent a human could.

As such, a disclaimer is shown pretty much everywhere, because you shouldn't take what the AI says as fact or accurate. However, it's great as a tool to aid you and learn more about yourself. Generally, the recommendation is that you share your journal entries and memory with a therapist who can help make sense of it all. We actually have a feature called `Medical Share` that's being developed so you can easily select your journal entries and securely share it with others. But more about that in a future update.

Currently, the Memory generation uses LLAMA 3.3 70B which is probably one of the best models we could've chosen. It's fast, works well with JSON structures, is open source and has enough parameters to give detailed answers. To be completely honest, we would have used DeepSeek R1 as our main AI Model considering that the reasoning could give even better insights but it's slower, has difficultly running tools and would've been almost double the price for a marginal improvement.

Privacy Concerns

Privacy is something that crucial to get right. Imagine if your data was brokered to ad companies or to other businesses. Or even trusting an AI with your private thoughts is a little scary.

So, we took a couple steps.

Firstly, we picked providers that don't store data or use your data to train models. Groq is where landed, within there privacy policy, they stand by not storing data at all and simply offering fast inference. We do also have an OpenAI Model (aka, the "Reliable Model" within the AI Chat) but we don't recommend it. OpenAI has done some pretty dodgy things.

Secondly, we encrypt the memory. The memory generated by the AI is created and stored in the same server meaning the memory is 256-bit AES encrypted at rest and transit.

Thirdly, we explicitly tell the AI not to store sensitive information such as address, phone numbers, emails etc. The AI returns a JSON object so it shouldn't even be possible to store such data.

Fourthly, I'm not profit hungry. Jadebook was built by 1 person, me, and so as long the App isn't losing money, I'm not interested in selling data. I've got no VC funding, no higher up I have to listen to or anything. The privacy policy fully states that we do not ever share your information and I stand by that. How many companies can say that?

Trying Out Memory

All free users get 20 credits for using AI. This should be enough to memorise a couple journal entries and ask a couple questions. You can also see what the AI generated.

As AI models and infrastructure improve, our AI features will undoubtedly get better and faster.

So try it out and feel free to share your feedback on improving JadeBook. We're currently in open beta and are working on fixing bugs and adding essentially, so if there's anything you can should be added, now's a great time to let us know.