Saving

What happens when you save a memory

The same steps run every time, in order, with no AI model involved.

Step by step

• 1. Check the input. Text must be non-empty and at most 20,000 characters, the space name 1–200 characters, at most 50 tags. Bad input is refused, never quietly fixed.
• 2. Look for hidden instructions in the text, the tags and the source, including tricks like look-alike letters, invisible characters and simple encodings.
• 3. Decide the status. Anything that looks like an instruction is quarantined. Everything else is active.
• 4. Check for contradictions with recent memories in the same space, and flag them for a person to review.
• 5. Label where it came from. If you say where the fact came from, that is recorded. If you don't, it is marked unverified. A label is not a trust score.
• 6. Fingerprint the text (SHA-256), so anyone can tell later whether it changed.
• 7. Save it and log it together. The memory and its log entry are saved in one step, so you can never get one without the other.
• 8. Link the log entry to the previous one, so the whole history forms a chain.

Why the check happens when saving

A poisoned memory can be read many times. Checking it once, at the door, means it is marked for good, instead of hoping every later reader remembers to filter it.

Example with a source and tags

curl -X POST https://cainstudio.online/aether33/api/memories \
  -H "Content-Type: application/json" \
  -d '{"space":"runbooks","text":"Rotate the signing key every 90 days.",
       "source":"security-policy section 3.2","tags":["keys"]}'