Your model bill stays yours. We never touch it.
You bring your own provider key and they bill you directly, at their price, with no markup from us. What Varve charges for is the memory: the record, the supersession, the per-person scoping. What it does to your model bill is order the prompt so the provider's cache actually hits — and then show you exactly what that was worth, in money, rather than asking you to take it on faith.
The arithmetic
your numbers, our ratesMeasured on this deployment: 35% on a cold prefix, 72% once warm. Across every paid call made so far — cold ones included — 26% of prompt tokens were served from cache.
Both columns are the same prompt — the same memory, the same tokens, sent to the same model. The only difference is whether the provider recognised the prefix. This is not a comparison against having no memory at all: memory costs tokens, and pretending otherwise would make the number meaningless. Rates are provider list prices in USD, carried from their published pages and not yet re-verified; your invoice comes from your provider, not from us. The Varve fee below is separate, in EUR, and we deliberately do not net the two off — that would need an exchange rate we would have to make up.
Run it with a cheap model and the saving is cents a month. That is the true answer and we would rather print it than hide the slider: caching does not pay for your subscription, and any vendor whose calculator says otherwise has chosen its defaults carefully. The saving becomes real money on frontier models and at volume — and the reason to buy Varve is the memory, not the rebate.
Plans
EUR, monthly, cancel anytimeOne team, one workspace.
| People remembered | 5,000 facts |
| Messages ingested | 20,000/mo |
| Assistant calls | 10,000/mo |
Several channels and a real backlog of history.
| People remembered | 50,000 facts |
| Messages ingested | 200,000/mo |
| Assistant calls | 100,000/mo |
Company-wide, with a dedicated database on request.
| People remembered | 250,000 facts |
| Messages ingested | 1,000,000/mo |
| Assistant calls | 500,000/mo |
Limits are enforced, not advisory — you are refused with a clear error rather than throttled silently or billed for an overage you did not agree to. A tenant that needs its own database rather than a row in a shared one is available on the business plan; ask and it is provisioned separately.
What you are not charged for
Your model spend
You bring your own provider key. We never proxy your billing, never mark it up, and never see your invoice.
Storage of your history
Every message is kept losslessly inside your fact allowance. We do not bill per gigabyte.
Leaving
Export is a documented endpoint, not a support ticket. Erasure destroys the key that makes a person's record readable and leaves the audit trail behind to prove it.
Talk to us
We onboard directly, a few teams at a time — there is no self-service signup, and provisioning is a person. Leave an address and what you run, and we will tell you honestly whether it fits.