Each app user's memory is their own.
Isolation by construction, erasure you can count, keys kept as digests and logs without bodies. What is not built yet is marked, not implied.
- IsolationEvery app user's memory kept apartLive
- Whose memoryFrom the key and a header, never a bodyLive
- ErasureEvery copy, with a receiptLive
- API keysSHA-256 digests; the token is shown onceLive
- Request logsMetadata only, kept 7 daysLive
- Encryption at restWith customer-managed keysSoon
- SOC 2 Type IIAn independent audit of these controlsSoon
Isolation
Every app user's memory is kept apart from every other's, so a request cannot reach across by mistake.
Whose memory a request reaches comes from your key and the X-Alexein-User header, never from a request body, so a model calling a tool cannot name another user.
| Key | X-Alexein-User | Memory |
|---|---|---|
| production key | user_42 | user_42 in production |
| production key | user_43 | user_43 in production |
| staging key | user_42 | user_42 in staging |
Erasure with a receipt
Forget hard-deletes every copy: the item as sent, the facts learned from it and everything learned from those, every version, and any copy still on its way in. It answers with a receipt: counts of what it erased.
A copy still on its way in, such as from a retry, cannot bring the item back. An item with an expiry is erased the same way, on time.
| Field | Count |
|---|---|
| items | 1 |
| facts | 1 |
| versions | 2 |
| waiting | 0 |
Keys are stored as digests
A key's token is shown once, when it is made. We keep its SHA-256 digest and its first ten characters, so the console can show which key is which without being able to show the key.
Revoking a key deletes what we store for it. A read key cannot write, and an MCP client given one sees no write tool.
{
"id": "key_3f9a1c2b7d4e",
"prefix": "tm_AbC1234",
"sha256": "9c1e5b7a…",
"org": "o7k2m9qx4ab",
"project": "production",
"label": "backend",
"writes": false,
"created": "2026-10-21T09:14:03Z"
}Request logs hold metadata, for 7 days
Each request leaves one line: when, which route, which project and app user, the status, the time it took and what it spent. No query, no item text, no response body. Lines are deleted after 7 days.
{
"at": "2026-10-21T09:15:42Z",
"request_id": "req_6b1f0e9d2a7c",
"route": "/v2/memory/search",
"project": "production",
"user": "user_42",
"status": 200,
"key_id": "key_3f9a1c2b7d4e"
}Where your users' text goes
Your users' text stays on our servers, except where a feature needs a model: then it goes to a model provider, through the subprocessors the DPA lists. The console's playground chat sends it too, but only when someone in your organisation uses it.
The service keeps no copy of a model's reply, since a reply holds your users' words. The playground chat is the one exception: it adds each reply to that app user's memory, with the message it answers.
| What | Where it goes |
|---|---|
| Your users' text | Our servers, and a model provider for a feature that needs one |
| The console's playground chat | A model provider, only when someone in your organisation uses it |
| A model's reply | Not kept, except a playground chat reply, added to that memory |
Reporting a vulnerability
Email security@alexein.ai with what you found and how to reproduce it.
- Test only against accounts and keys you own.
- Read no more of anyone else's data than it takes to show the issue, and tell us what you saw.
- Give us a reasonable time to fix it before you publish.
- Leave out denial of service, spam and social engineering of our staff.