The Amnesia Bottleneck

Ask why AI-assisted engineering stalls at scale and you get the loud answers: the model is not good enough, the context window is too small, the tooling is immature. Real, but not the whole story. There are four bottlenecks that show up once you push agents past a toy project, and one of them almost nobody names, because it hides inside every session that starts clean.

That one is amnesia.

Every session, your agent re-learns the codebase from scratch. It re-derives the architecture, re-discovers the conventions, and re-opens decisions you settled months ago. You spend the first twenty minutes of every session paying that tax again, and the agent still gets it wrong, because relearning from the code in front of it can only recover what the code shows, not why you chose it or what you have decided since. The bill compounds with every project you keep alive. The more history you accumulate, the more there is to forget.

The instinct is to fix it with a file of notes or a vector database. Both fail the same way, from opposite ends.

A static note file freezes your old facts and lets them dominate forever. Last year’s decision outranks this year’s judgment because it was written down first and never moves. A vector database freezes your understanding at the instant you stored it. The retrieval feels dynamic, but the belief underneath is set in amber. Feed either one to an agent working an old project and it will confidently rebuild the version of you who started it.

Memory that applies what changed since

Curator is the third thing. It is evolutionary memory, and the difference is that it applies what changed since you last touched a project instead of replaying the past unaltered.

Evidence accumulates. Each memory carries a sense of how well-established it is, a value that grows as corroboration arrives and decays with age, so a decision you keep reaffirming rises and a one-off status note sinks. Superseded decisions are held back rather than surfaced, so the standard you have since moved on from stops steering new work. When two memories contradict, recall resolves to the better-supported belief instead of the louder or newer one. And when nothing genuinely fits, recall refuses to hand back a weak match dressed up as an answer. It would rather return nothing than confabulate, because a confident wrong answer is the one that costs you the most.

The result is that returning to a project no longer starts from zero. You resume with what you know now: the decisions you have made since, the standards you raised in other work that should apply here, and the patterns this code still uses that you have since rejected.

The memory layer of the sovereign stack

Amnesia is one bottleneck of four, and the fix is not a bolt-on. It is a layer the rest of your stack feeds and reads. Curator is that layer in the sovereign AI engineering stack, and it stays on your terms: it runs local, encrypted at rest, keyed to your machine, with your own decisions held sovereign over anything a machine derived on its own.

It does not sit alone. The analytics that TheAuditor produces about your code flow in as ground truth, so memory is anchored to what the codebase actually is, not just what you remembered about it. And Warden surfaces that memory in a native tab, so the agent acting on your behalf has your current judgment already in front of it instead of relearning it mid-task.

Remove amnesia and every session stops restarting. That is the point. Your agents inherit the engineer you are today, not the one who opened the project a year ago.