I built a memory layer for AI agents because mine kept forgetting everything

For the last few months I kept hitting the same wall.

I work with AI agents every day. The hard part was not getting them to do the work. The hard part was that they forgot everything the moment a chat ended. I would explain my project on Monday. By Wednesday the agent was a blank slate again. So every morning started the same way: paste the context, explain the goals, repeat the decisions we already made. Then do real work. Then do it all again the next day.

It felt like working with a smart helper who had no memory at all.

So I tried the usual fixes. I saved my notes and let the agent search them. This helped a little, but it kept pulling back the wrong things. I would ask what a teammate said about a price change, and the agent would hand me three old files that just used the word "price" somewhere. Close, but useless. The agent was matching words, not understanding what actually happened.

That is when the real problem clicked for me. Memory is not a pile of text.

Memory is about how things connect. A single decision has people in it, a reason behind it, and a project it belongs to. Flat text search throws all of that away.

So I started building HyperMemory.

The core idea is a hypergraph. I know that sounds fancy, so here is the plain version. Picture a normal graph: a line that joins two dots. That can only say "A is linked to B." A hypergraph link can join three, four, or more dots with one line.

So instead of breaking a meeting into a dozen tiny pieces, I can store the whole thing as one memory: who was there, what they chose, and why. When the agent looks it up later, it gets the full story back, not random fragments.

Under the hood, when you save something, the system quietly pulls out the people, projects, and decisions and wires them together. Later it can answer real questions like "who was part of the Q3 plan and what were their concerns?" That is the kind of thing flat search could never do for me.

The part I am most happy with is how easy it is to try. There is no SDK to install. You paste one URL into your AI client (Claude, Claude Code, Cursor, Windsurf, and others all work), sign in, and your agent suddenly has memory tools. It took me far too long to get there, but now it really is about a minute of setup.
Now the honest part, because I know this crowd values it.

This is still early. The hypergraph adds a little delay when you save things, because it does that background work of linking everything up. And to be real with you: if you only need to remember one simple fact, this is overkill. A plain note would do the job. HyperMemory earns its place when your questions get messy and involve a lot of moving parts and people. That is the exact spot where I was losing hours every week.

A few details for anyone who wants to poke at it:

The free tier is 3,000 queries and 3,000 nodes a month, with no credit card.

Reads count against queries; saving facts does not. Paid plans add more room and more queries, and that is how I plan to keep the lights on. Your data is hosted in the EU and is never used to train any model. You can also export it whenever you want, so you are never locked in.

Here is my one real ask. Does the hypergraph idea actually land for you, or does it sound like I over-engineered a simple problem? I have been staring at this for months, so I have lost the ability to tell. I would rather hear a hard "this is too much" now than find out later.

If you want to try it, it is free to start here: https://hypermemory.io

Thanks for reading. I will be in the comments all day and I will answer everything.