Answers · a straight answer, sourced
How do I give a local LLM long-term memory?
A local model runner serves tokens and forgets everything between sessions, so long-term memory has to come from a layer around it: something that stores durable facts, retrieves the relevant ones into each prompt, and — if the memory is going to be trusted — records how each fact was learned.
Most "memory" tutorials tell you to paste facts into a `MEMORY.md` file and prepend it to every prompt. That works until the file is large, contradictory, or you want to know why the model believes something. Real memory needs three things a text file does not have: extraction (turning conversations into durable facts), retrieval (pulling only the relevant facts into context, so you are not paying for the whole file every turn), and provenance (which conversation a fact came from, and which model wrote it).
Provenance is the part almost nothing does, and it is the part that makes memory safe to rely on. If your assistant "remembers" that you are allergic to a medication, you want to know whether that came from a document you imported or a model’s guess — and to strike it, permanently, if it is wrong.
SovereignAI is built as exactly this layer: durable memory where every fact records its origin (typed, extracted, or distilled), the source conversation, and the authoring model; retrieval that puts the best passage first; and deletion that zeroes the bytes on disk rather than hiding a row. It sits in front of whatever generates the tokens — Ollama, a frontier API, or a local sparse model — so the memory is yours even when the brain is rented. One honest caveat: extraction quality still depends on the model doing the extracting; a small local model writes rougher memory than a large one.
Own the answer, not just read it
SovereignAI is private AI you run yourself: local models, memory with receipts, disclosure at the moment anything leaves, and an exit you can verify. The trial is one Docker command, and deleting the volume deletes every trace of you.