Why this policy exists
Most of what Kemory does costs us almost nothing per user. Capturing your conversations, storing them as memory, searching them, and injecting context back into ChatGPT or Claude all run against your AI subscriptions and your storage allowance. Those are governed by your plan limits, not by this policy.
Two features are different. Ask Kora and Continue with Kora run inference that we pay for. Every time you ask Kora a question, or carry a conversation into Kora from another AI, we send tokens to a model on your behalf and pick up the bill.
We would rather publish honest limits than advertise "unlimited" and quietly throttle you. This policy states what generous individual use looks like, what happens if you go past it, and what we will never do to your data in response.
What this policy covers
| Feature | What it is | Metered here? |
|---|---|---|
| Ask Kora | You ask a question; Kora routes it to an appropriate model, grounded in your memory | Yes |
| Continue with Kora | You take a conversation started in another AI and continue it inside Kora, with your context carried across | Yes |
| Counsel | You put one question to several models at once and receive a synthesised answer | Yes — see below |
| Capture and sync of conversations | Reading your chats from connected AIs into your memory | No — plan limits apply |
| Memory storage and search | Your namespace, retrieval, dialectic synthesis | No — plan limits apply |
| Continue in another AI | Moving a conversation from Kemory into ChatGPT, Claude, Gemini and others | No — runs on your subscription |
| MCP tool calls from your own agents | Reads and writes against your namespace over MCP | No — API rate limits apply |
Units, and what counts
One Ask Kora request is one question and its answer, including any memory retrieval and any follow-up tool calls Kora makes to answer it. A clarifying question you ask in reply is a new request.
One Continue with Kora handoff is one conversation carried into Kora. The handoff includes compaction of the source conversation and the first Kora turn. Every turn after that is an Ask Kora request.
One Counsel query is one question fanned out to multiple models plus the synthesis step. Counsel is counted separately because it costs roughly as much as the number of models convened.
What does not count against your allowance
- Requests that fail because of an error on our side (5xx), or that we retry ourselves
- Requests rejected for authentication or permission reasons before any model is called
- An identical request repeated within 60 seconds where we serve a cached answer
- Anything you cancel before the first token is returned
Allowances
We are not publishing numeric allowances yet. Kemory is in its first release and we are still calibrating what generous individual use actually looks like against real usage. Publishing thresholds now would mean picking numbers we would very likely have to change, and we would rather set them from evidence than from a guess.
What we can tell you is exactly what we measure. Usage on the metered features is counted per account along these dimensions:
| Ask Kora | Requests in a rolling window, and requests per day. |
| Continue with Kora | Handoffs per day, and the size of the compacted context carried across in each one. |
| Counsel | Queries per day, and how many models each one convenes. |
| Concurrency | How many Kora requests you have running at once. |
| Cost | Total inference cost per account per calendar month, as a backstop for patterns the counters above do not anticipate. |
Rolling windows are rolling, not fixed — you are never waiting for a clock at the top of the hour. Daily counters reset at 00:00 UTC. Team allowances pool across the subscription rather than being enforced per seat; we do not police per-person distribution inside a team. Enterprise limits are whatever your contract says.
When we set thresholds, they will be published here with a version and a date, and account owners on paid plans will be emailed before any of them start to bite. Until then, the escalation ladder in the next section is what governs — and its first step is always telling you, not restricting you.
What happens when you go past an allowance
We escalate in order, and we tell you at every step.
| Approaching a limit | The extension shows a quiet indicator. Nothing changes. |
| At the limit | We may restrict or degrade the metered features rather than stopping you outright — typically by routing Ask Kora to a lighter model, and by pausing Continue with Kora and Counsel until the window resets. Where a request is answered by a different model, you will be able to see which one. |
| Sustained overage | Several consecutive days at the limit — we may reduce concurrency and queue requests behind other users during peak hours. |
| Persistent overage | We email you. The conversation is about finding the right plan, not about cutting you off. For a heavy individual user that usually means a team plan; for a heavy team account, Enterprise. |
| Abuse | Under prohibited use, we may throttle, suspend Kora features, or terminate access without notice. |
We will not delete, hide, downgrade, or refuse to export your memories because of a usage dispute. Your memory is yours. Capture, search, export and continuing into another AI keep working even if Kora features are paused. This is not negotiable and it survives any future version of this policy.
Prohibited use
The following are outside fair use on every plan, including Enterprise unless your contract says otherwise:
| Automation | Driving Ask Kora, Continue with Kora or Counsel with scripts, bots, headless browsers, or scheduled jobs. These features are for a person at a keyboard. Programmatic access belongs on the API, under its own rate limits. |
| Account sharing | More than one person on a single Free or Pro seat. Team exists for teams. |
| Reselling or proxying | Exposing Kora as an inference endpoint inside your own product, or reselling access. |
| Training and distillation | Using Kora outputs to train, fine-tune, evaluate or distil a competing model or memory service, or systematically harvesting responses for a dataset. |
| Circumvention | Multiple free accounts to extend allowances, rotating identities, or interfering with metering. |
| Load that degrades service | Any pattern that materially affects availability or latency for other users. |
Enforcement under this section may be immediate and does not require prior notice.
Model routing
Ask Kora chooses a model per question. Light questions do not go to frontier models, which is how we keep Kora viable at its price. Under sustained load, or once you are past an allowance, routing biases further toward open-weight models we run ourselves.
You will always be able to see which model answered. If you need a specific model for a specific question, use Continue in another AI and run it on your own subscription — your memory travels with you either way.
Changes to this policy
This policy is subject to change. Kemory is in its first release and the underlying model prices move. This version sets out what we meter and how we escalate; the thresholds themselves are still being calibrated and will be published here when they are settled.
- Every version is numbered and dated, with a changelog at the foot of this page.
- Material reductions to allowances on paid plans take effect no earlier than 30 days after we notify account owners by email.
- Increases, clarifications, and changes to Free plan allowances may take effect immediately.
- Changes made in response to abuse, security incidents, or provider-side changes outside our control may take effect immediately on any plan.
- If a material reduction lands mid-term on an annual plan and you would rather not continue, contact us and we will refund the unused portion.
Continued use after a change takes effect is acceptance of the revised policy.
Questions and appeals
If you have been throttled or contacted and you think we have it wrong, reply to the email or write to support@sekondbrain.ai. We would rather move you to the right plan than lose you.
If you are consistently at the ceiling and the next tier does not fit, tell us — an unusual usage pattern is usually a product signal, not a violation.
Changelog
| Version | Date | Change |
|---|---|---|
| 1.0 | 12 August 2026 | Initial publication. |