#The Problem with AI Memory (and How We Fixed it)*

This morning, I was emailing with an AI YouTuber in response to a question about what makes Basic Memory different from Chat’s memory system and what we presume will be Claude’s Memory, if and when they get around to releasing it. I found myself writing at such length on the topic, I thought it might be a good idea to write more about it here.

Note: They did, read our take here.

Anyone learning about Basic Memory for the first time might say, “Chat already remembers stuff about me.” And, certainly, that’s true. It does remember some things. But not in a way that belongs to you or delivers the highest value.

Let’s look at a few examples of conversations that are impossible with Chat’s memory system that are a breeze for Basic Memory.

If you look at Chat’s memory, you’ll see things like “Drew lives in Austin” or “Drew hates tomatoes.” A really beefy memory in Chat’s database might explain a specific project or even some details about it. Some of those memories are even helpful, but their inevitable incompleteness can also work against you. Given limited information, LLMs begin filling in the blanks with what “feels” to it like the logical middle, even if it’s completely incorrect. They start making weird, incorrect assumptions. But what can you do? You can’t edit memories in Chat. You can only delete them. If, that is, you can find them in a lengthy scroll of (unsearchable) memories.

Basic Memory takes a drastically different approach. I like to use the image of a stenographer sitting in on all your chats and taking detailed notes on all your interactions. The notes are finished the moment the conversation ends, written in plain text you and your LLM can both read, search, and edit. You can also write notes or add context independently that your LLM will use thereafter. It’s all part of your second brain that grows as you do.

Every note becomes part of a semantic knowledge graph that discovers connections that prove invaluable over the life of any project, even if that project is just day-to-day living. These aren’t bullet points or lists. They’re fleshed out notes that make sense of what you’re trying to achieve and where things stand, not merely as “facts,” but as entries in a narrative that also take temporality into account.

With Basic Memory, you can understand when you arrived at a decision and by what logic. It’s not separated into two talking points like “Drew is debating buying the Japanese movie poster for Pump Up the Volume,” and “Drew is the proud owner of a Japanese movie poster for Pump up the Volume.” Your notes hold all the data in between, making sense of the trajectory of thought and experience that led you to the present circumstance.

That’s not all that gets lost in a typical LLM chat. In my experience, and the experience of several of our users, chats with AI have a quality that might best be described as “shaggy.” That is to say, they’re meandering, unpredictable, and spontaneous. You start off talking about a professional project, briefly ruminate about an ex, fantasize about a business idea, then return to that professional project. If you want to revisit that brilliant idea sandwiched between other things, you might never find it again. But with Basic Memory, conversations don’t have to be tidy or neat to be remembered. Like a kind of informational stew, everything goes in.

But all of that is merely a precursor to a linchpin notion that is central to the entire Basic Memory project and ethos:

All of it. And not buried in json files you can request from a vendor or hidden away in vector databases. These are your chats, your ideas, and your discoveries. They should be accessible, truly readable, and under your control. That’s what Basic Memory makes possible. Reddit is filled with stories of people who find themselves inexplicably barred from Chat or Claude. Usually (at least in the cases I’ve seen), it’s an error. Do they ever recover their chats, who knows? But it seems crazy to surrender that much power and trust to entities whose customer service reps are notorious for being all but unreachable no matter how desperately you try. For many of us, our AI chats are where our ideas now live. What’s more important than protecting that?

So, again, how is Basic Memory different than an LLM’s memory? In just about every way. It’s exhaustive instead of reductionistic, temporal instead of based on static facts, editable instead of locked down. It’s searchable, it draws connections you might never have drawn, you can write and edit your own “memories,” and, most importantly, it’s YOURS.

If you check out our Discord, you’ll see our users utilizing Basic Memory in such wildly unexpected ways—from a guy using it to make music, to coaches using it to track the progress of their clients. We recently talked to one user making a (quite complex sounding) dream tracking system and another sharing his Basic Memory vault with a team of 6 to 7 people as they carry out deep data research.

As we prepare to launch Basic Memory Cloud, our thoughts are on all these issues, in the hopes that Cloud will open up even more possibilities to our users.

If you haven’t tried Basic Memory yet, we’ve just released a very straightforward one-step install process. You can find our videos about it here: https://basicmemory.com/videos/

And, truly, check out our Discord. It’s filled with inspiring use cases and some fascinating characters doing things we never would have imagined.

*Obvious Clickbait Headline: AI Memory is an unsolved problem, we are working on it.