Reference

How AI features work

Search by meaning, descriptions and tags, personal keys, allowances, and what leaves memex.

Settings › AI features has two sections. Each shows what it does, its usage allowance, and the provider key it uses.

Search

Memex prepares new and edited notes so it can find them by topic as well as exact words. Two counters show today’s remaining allowance for preparing notes and for searches beyond exact words, or say there is no daily limit. Exact-word searches do not consume the search allowance and still work when it runs out. The limits may change.

Adding a personal OpenAI API key removes Memex’s limits; OpenAI bills you for usage and its own limits still apply. An OpenAI key added under Descriptions and tags is listed here too but not used for search until you press Use for search on it. On memex.tools the embedding model cannot be changed.

Descriptions and tags

When enabled, Memex writes a description for a note saved on the website that needs one and adds relevant tags from the existing vocabulary. It never describes or tags what an assistant writes; the assistant does that itself. Supplied descriptions and chosen tags are kept. New tag names are suggestions only; system tags are excluded.

The section shows On or Off. When on, it shows the remaining daily allowance or says there is no daily limit. Where it is already on, it names the model that writes the descriptions and shows the current allowance; otherwise it needs a key of your own.

Adding a personal key from OpenAI, Anthropic or Google enables this feature and removes Memex’s daily limit. The provider bills you for usage. You can choose a model and turn Add descriptions and tags automatically on or off, saving those changes. A small, fast model is enough. The editor’s title and summary tools use the same key.

Note maintenance

Note maintenance is in Settings › Assistants. It runs on your assistant’s account, not on a provider key stored here. See How curation works.

Keys

Add personal key asks for a name, the provider and the key. memex makes one real call to the provider before storing it, so a key that does not work is never saved. It is stored encrypted and shown afterwards only as its last four characters. Delete stops anything set to run on it, and that section goes back to its allowance.

Allowances

Search and descriptions are limited per knowledge base, by hour and by day. The allowances are generous and may change. When one is reached, a save that needs a new embedding is refused with a message saying when it will work again, and nothing is lost; search never refuses, it falls back to keyword matching and says so. A section running on your own key has no allowance at all.

What leaves memex

Note text goes to OpenAI for embeddings. With descriptions on, the note’s text, title and your tag vocabulary go to OpenAI, or to the provider of your own key. The editor’s duplicate, link and tag checks run on your own data and send nothing anywhere. memex does not train on your notes; connected assistants have their own terms.