Privacy and data handling
Inferam is built to keep your code and prompts private. The binding text is the privacy policy; this page explains how it works in practice.
What we never store
Section titled “What we never store”- Prompts, messages, system prompts and tool results you send
- Completions, reasoning and tool calls the model returns
- Images or other media you send
Inferam never writes these to its databases, logs, traces, analytics or error reports. Error handling records only which route failed and the error’s type, never the message, which could quote your input. The one place request content persists is the model server’s prompt cache, described below.
We never train on your data, nor on anything the model returns to you.
What we store for each request
Section titled “What we store for each request”Exactly this metadata, and nothing else:
| Field | Why |
|---|---|
| Account ID and API key ID | Your usage dashboard, and per-key usage |
| Model and timestamp | Usage over time |
| Input, output and cached token counts | Usage totals and graphs |
| Time to first token and generation time | Tokens per second, and capacity planning |
| Status code | Error rates and reliability |
This metadata is deleted after 6 months, by a job that runs daily. You can see your own on inferam.ai/usage. Inferam also keeps running lifetime totals of input, output and cached tokens and generation time across all accounts, with no per-account figures, and publishes them as public stats.
Inferam also stores your email address, your API keys as irreversible SHA-256 hashes (the full key is only ever shown to you once), and your web sign-in sessions. You can end every web session at once with Log out everywhere on inferam.ai/keys.
Prompt caching
Section titled “Prompt caching”Coding agents resend the whole conversation with every request. So the model server doesn’t recompute it each time, it keeps the processed state of recent prompts, and a later request that starts the same way reuses it.
- Where: in memory and on the local disks of the model servers on Inferam’s own hardware. It never leaves those machines and is not copied to Inferam’s API, databases, logs, monitoring or backups.
- What: the model’s internal state for your prompt, together with the prompt’s tokens. Treat it as containing your prompt, including the earlier replies your agent resends as conversation history.
- How long: deleted automatically no later than 7 days after it is written, often sooner: the cache is capped at 64 GiB per server and the oldest entries go first. It survives server restarts.
- Encryption: the cache disks are not encrypted at rest. The servers are not reachable from the internet.
Where requests go
Section titled “Where requests go”- Your request reaches Inferam’s API on Cloudflare over HTTPS.
- The API checks your key, waits for a free slot, and forwards the request over a private, outbound-only tunnel to the model server on Inferam’s own hardware. The model server isn’t reachable from the internet.
- The response streams back the same way.
Operational monitoring (Grafana Cloud) receives metrics and the request metadata above, never request content. We check this with a canary test: a unique marker string is sent through every API format, then the logs, database, model-server logs and monitoring are searched for it. Any hit is treated as a privacy bug.
Coding agents
Section titled “Coding agents”The inferam CLI configures your agent for one session and doesn’t change its settings. For DeepSeek Harness, it also turns off the session-log upload and telemetry that would otherwise send session content to DeepSeek. Other agents’ own telemetry, which goes to their vendors, is controlled by each agent’s settings.