AetherMemAetherMem

Every model I used had the same problem. Somewhere in a long conversation it would lose the thread, and reminding it never quite fixed that. It would contradict a decision it made an hour earlier, forget a rule I set at the start, or chase the last thing I said instead of the thing that mattered. I got tired of babysitting the context window, so I stopped patching my prompts and started building the model a memory that manages itself. Making the window bigger doesn't fix the problem. The model needs a better way to decide what to keep.

A retrieval layer scores all four by importance and blends them whenever the model needs to remember something, so it recalls what matters instead of whatever it read last. On top of that sits a diagnosis engine: a second model that reads the system's own results and failures and traces each one back to a cause. When something goes wrong, it works out why.

The part I'm proudest of is the loop that keeps it honest. It proposes a change to itself, applies it, checks whether things actually got better, and undoes the change if they didn't. Letting a memory change itself is easy. Stopping it from changing in the wrong direction is the hard part. Everything it does is written up in plain language on a dashboard called the Chronicle, so you can read what it chose to remember, what it chose to forget, and why.

It runs end to end today, and I'll call it what it is: a research prototype, and not a finished product. I'm still building it. Next, it learns to write tests for its own failures and to reason about how it scores itself. It's the project I'm deepest into right now, and the one that keeps me up at night.