An agent without memory is just a clever stranger

The most capable model on the market still wakes up with amnesia. Open a new session and the context window is blank — no memory of the decision you made last week, the gotcha that cost you an afternoon, or why you deliberately chose the boring option. The model is brilliant for the length of one conversation, then forgets all of it.

That’s the single biggest gap between a demo and a desk you actually run on.

The amnesia problem

Without persistent memory, every session re-establishes the same ground: which stack this project uses, what you prefer, what’s already been tried and ruled out. You become the agent’s memory — repeating context, re-explaining choices, catching the same mistake twice. It works, but it doesn’t compound.

Satori-Kura

This desk runs on a self-hosted, private memory layer called Satori-Kura. It isn’t a chat log. It’s a structured, searchable knowledge base the agents read at the start of work and write to as they learn. Each memory is a standalone statement — written so it makes sense with zero conversation context — tagged with a type (fact, decision, lesson, preference, project) and scoped to a project so unrelated work never bleeds together.

Storing text is the easy part. What makes it cognitive is the behavior around the text.

The features that matter

Hybrid recall — semantic (vector) search fused with keyword search, so a query finds the right memory whether you remember the exact words or just the gist.
Project isolation — memory is scoped per project. The client site’s quirks don’t surface while writing about infrastructure, and vice versa.
Corrections, not deletions — when something changes, the old memory is retracted and superseded, preserving an audit trail of how the understanding evolved — not just the current answer.
Private & self-hosted — the whole layer runs on owned infrastructure, not a third-party service. The desk’s accumulated knowledge — clients, decisions, lessons — stays under one roof: nothing leaves for someone else’s model to train on, there’s no per-seat SaaS meter, and no vendor can change the terms, raise the price, or pull the plug. Memory this valuable shouldn’t live on rented land.

Two more features matter less for everyday recall and more for the moment something goes wrong — when a remembered “fact” looks off, or you need to reconstruct why the agent did what it did. This is the audit trail.

Confidence labels — how much to trust a memory

Not every memory deserves equal weight, so each one carries a one-word confidence label that rides along on recall:

Label	What it means
`observed`	The default — something the desk recorded firsthand.
`inferred`	The agent worked it out itself, without direct evidence. Handle with a little more care.
`user_confirmed`	A human acknowledged it — a logged lesson or a decision.
`authoritative`	Human-authored, foundational, and shielded from decay. The bedrock.
`stale`	Flagged as likely out of date, so it’s downweighted when results are ranked.
`rejected`	Retracted — kept only as a record of what was once believed.

A nightly housekeeping pass re-derives these labels from each memory’s current state, so trust stays in sync with reality without anyone hand-tagging it.

Together, the confidence labels and that paper trail are what make a memory issue diagnosable. When a recall looks wrong, you can see how sure the system was, where the fact came from, and what put it there — instead of guessing.

Why it compounds

Each of these is small on its own. Together they change the economics of working with an agent. Corrections stick. Decisions stay decided. A lesson learned the hard way on one project is there, unprompted, the next time its shape recurs. The desk doesn’t just retain information — it gets sharper the longer it runs, because every session writes back what it learned for the next one to stand on.

That’s the part you can’t fake with a bigger context window. Context is what an agent is holding right now; memory is everything it has earned and kept.