Non-parametric memory
Knowledge a model pulls in from outside itself at answer time, rather than from its weights.
Think of it like
Like an open-book exam — the facts live in the book on your desk, not only in your memory.
Example
A RAG system fetches the exact policy paragraph and hands it to the model, so the answer comes from a document, not the weights.
How it actually works
Non-parametric memory keeps knowledge in an external, swappable store — documents, a vector database, a tool — that the model reads on demand. Its advantages are updatability and citation: you can change a fact without retraining, and the model can point to where an answer came from. The cost is a retrieval step that can fail.
For product teams
How you keep answers current and auditable without retraining the whole model.
For engineers
Knowledge held in an external store queried at inference; updatable and attributable, unlike weights.
Related
- Parametric knowledge — Its in-the-weights counterpart.
- RAG — The retrieval pattern that uses it.
- Vector database — A common backing store.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome