One price.
MemRails owns the hard part — evidence-graded retrieval, the packet contract, the network. The model that writes the packet is your choice: bring your own key and pay your provider directly, or let us run one for you. Either way, MemRails charges a thin orchestration fee for the retrieval and the contract — and nothing for the inference you didn't ask us to run.
The retrieval, evidence assembly, and contract enforcement that turn your corpus into a packet-ready bundle. Thin, flat, and identical no matter which model you point at it.
Route the packet write-out to any model. BYO key and pay your provider directly — MemRails sees none of it — or pick our managed Compress-v1 and we bill the tokens. The brain is swappable; the contract isn't.
Bring your own, or let us run one.
You hold the model relationship.
Connect your provider key — frontier, small, or self-hosted. MemRails assembles the evidence bundle and hands it to your model under the packet contract; the tokens bill to your provider, never to us.
- +MemRails inference cost to you: $0
- +You pay your provider their token rate, on your existing contract
- +MemRails charges the orchestration fee only
- +Swap models any time without touching your MemRails bill
We run the default, tuned for the contract.
Don't want to manage a model? Use Compress-v1 — our model fine-tuned specifically for the packet contract, served on our GPUs. We bill its tokens at a managed rate; you bring nothing.
- +No key, no infra, no model ops on your side
- +Tuned for packet fidelity — citations, contradictions, provenance
- +Token rate billed by MemRails, marked over GPU cost
- +Still swappable — start managed, move to BYO whenever
Own the memory. Rent the brain anywhere.
No per-seat tax.
Add a thousand engineers or a thousand agents; your access cost does not move. You are billed for queries run, never for headcount. Procurement models one line, not a seat-license true-up every quarter.
Predictable to the dollar.
Finance can forecast spend from query volume with no step functions, no tier cliffs, and no surprise overage classes. What the calculator shows is what the invoice says.
No reason to lock you in.
We are not incentivized to lock you in — we are incentivized to make the memory worth querying. Bring any model, leave any time; the only thing keeping you is whether the packets are good.
Auditable by design.
Every metered query and every settled fee is logged with a timestamp and an immutable reference. Reconcile our invoice against your own telemetry line by line.
Swap the model. Watch our fee hold still.
Pick a compression model and set your volume. The MemRails line — the orchestration fee — doesn't move when you change models. Only the compression line does, and it lands on our bill or your provider's depending on how you run it.
The questions a buyer asks.
The compressed answer an agent consumes at the end of a query. MemRails gathers the candidate evidence across L1–L4, and a model synthesizes it into one structured bundle — typically ~500 output tokens — carrying provenance, surfaced contradictions, and calibrated confidence. It's the billable unit: one packet is one completed, evidence-graded synthesis, not a raw document dump.
One thing: the $5 / 10K packets orchestration fee. The compression tokens bill to your provider on your own contract — MemRails never touches that inference cost, and there are no seat fees on top.
Everything that turns your corpus into a packet-ready bundle: retrieval across L1–L3, evidence filtering, the structured packet contract, provenance tracking, and network access. It's the work that decides which evidence is worth compressing — which is the part that's genuinely hard to reproduce.
The contract is enforced regardless — every packet carries provenance, surfaced contradictions, and calibrated confidence. A stronger model fills it more faithfully; our managed Compress-v1 is tuned specifically for it. You trade off model cost against fidelity, with the floor held by the contract.
To skip model ops entirely. Compress-v1 is fine-tuned for the packet contract and served on our GPUs — no key to manage, no infra to run, and predictable token pricing. Start there; move to BYO the day you want to.
Yes — it's priced on packets, not tokens, so it doesn't move when you switch models. The calculator makes this visible: change the model and the MemRails line holds while only the compression line shifts.
Yes. Committed-use lowers the per-packet orchestration rate and fixes it for the term. If you run managed Compress-v1, the token rate is committed too. BYO inference stays on your provider's own commitment.
None. Pricing is on packets and the tokens they consume — never on headcount. Add a thousand engineers or a thousand agents and the access cost doesn't move; you're billed only for the queries that actually produce packets.
Both leave with you. Markdown and Git history export in one operation; usage and settlement logs export as JSONL. Your model routing was always your own key against your own provider — there's nothing of yours locked inside MemRails to reclaim.
Modeled before you sign.
Bring your projected packet volume and model choice. We'll turn the calculator above into a committed-use proposal — fixed rates, audit terms, and an SLA in writing.