Why your local model forgets you — and how to give it a memory you own

You run an open-weights model under Ollama, LM Studio or llama.cpp because you want control — over your data, your costs, your stack. The model is capable. And yet every session starts the same way: it has no idea who you are. You re-explain your projects, your preferences, the way you like answers shaped, and tomorrow you do it again. Here’s why your local model forgets you — and why that’s the easiest forgetting of all to fix on your own terms.

Why local models forget

It isn’t a flaw. It’s how the models are built.

A language model is stateless. The weights you downloaded encode general capability — language, reasoning, code — but they hold nothing about you. Within a single conversation, the model only “remembers” what’s inside its context window, the span of text it can see at once. When the session ends, or the window fills and older turns scroll out, that context is simply gone. There’s no notebook the model writes to between chats.

Cloud assistants paper over this with a memory feature: a service alongside the model that quietly saves facts about you and feeds them back in. Your local setup has the model — but not that surrounding service. So it starts blank every time. The forgetting isn’t the model failing; it’s the absence of a memory layer around it.

Why this is good news

The instinct is to envy the cloud assistants their memory. Resist it. The reason your local model has no memory is the same reason it can have the best possible one: there’s no vendor in the middle.

Cloud memory is convenient but locked in — it lives on the vendor’s servers, in the vendor’s format, and only works inside that one app. Your local model has none of that baggage. Whatever memory you give it can be a plain file you hold, on your machine, that never has to leave. The gap in your local setup is exactly the place an AI memory you own slots in.

How to give your local model a memory you own

The fix is a memory layer: a persistent store of who you are and how your AI behaves, sitting beside the model and feeding it the right context each session. Two parts matter.

Your persona — the AI’s character with you. Its voice, the conventions you’ve agreed, the things it should refuse, the way it greets you. This is what turns a generic model into your assistant.

Your corpus — what the AI knows about you. Your projects, the people in your life, your preferences, the facts you keep having to restate. A local, searchable store the model can draw on.

With aiperson, both live where a local-model user wants them:

• A signed file you own. Your persona is kept in a git repo you control; your corpus is a local SQLite database on your machine.
• Nothing hosted by default. This is the point for a local setup. Your memory need never touch anyone’s servers — it’s local-first, the same way your model is.
• Projected into the model each session. The persona is read and put in front of your local model so it opens already knowing you, instead of from zero.
• It learns as you work. Rather than you maintaining a notes file by hand, the layer records what’s emerged worth keeping, and your persona grows in your own repo.

The result: your Llama or Mistral under Ollama greets you by name, holds the conventions you set last week, and remembers what you were building — with nothing leaving your machine.

The proof that the memory is really yours

Here’s the test that settles it. Cultivate an assistant under a cloud model like Claude for a week — let it learn your voice and your projects. Then point your local Llama at the same persona file. If it greets you the same way, keeps the same refusals and remembers what you were working on, that continuity is the thing you own.

That move — cultivate under Claude, then run it on a local model, persona intact — is the un-fakeable proof. No single vendor can offer it, because it spans all of them. And it works precisely because the memory is a file you hold, not a feature you rent.

You already chose ownership — finish the job

Running a local model is already a decision to own your stack. Giving that model a memory you own is the missing half of the same decision. Your weights are yours; your memory should be too.

Local models aren’t an afterthought here — they’re a first-class audience. If you want to go deeper, see how to give your local AI a memory that belongs to you, or own your AI memory today.