Your AI assistant remembers. That is the point of it — it recalls last quarter’s decisions, the client’s history, the case that is still open. What most organisations have not yet asked is the harder question: where does that memory live, and who else can read it?
For anyone handling regulated data, the answer matters more than the convenience. Once a model remembers a patient, a claimant, or a customer, that record is subject to the same obligations as any other copy you hold. You need to say who can read it. You need to erase it on request and show that you did. In most AI tools today you can do neither, because the memory lives inside someone else’s service.
Memory you can actually account for
SAIHM inverts that. Memory is sealed on your own machine under keys you hold. The operator stores only ciphertext it cannot open — not as a policy promise, but because it never has the key. Erasure destroys the key and leaves a receipt you can hand an auditor. Sharing a record with another agent is explicit and revocable, not a blanket handover.
What changed this week
The compliance story was already true. Getting started was not.
Until now, activating a memory of your own meant generating a 64-character master secret by hand and passing it to your agent before anything worked. That is a reasonable ask of a security engineer and an unreasonable one for everybody else — and an autonomous agent could not do it for itself at all. It was the single step where most people stopped.
That step is gone. You ask your agent to join, and it generates and seals its own identity on your machine. No key handling, no card.
Join SAIHM, then remember: our retention policy is 90 days
for support transcripts, and Q3 renewals are handled by the
named account team only.
Later, in a different session — or a different model entirely:
Recall what you know about our retention policy.
The memory is still there, still sealed under your key, and still erasable on demand with proof.
Making it a habit
Most of the value comes from one standing instruction, given once. A memory contract tells your assistant to check what it already knows before it answers, and to keep anything that will matter later:
Memory contract: at the start of every session, recall what
you know about me from SAIHM before answering. When I tell
you something that will matter later, remember it.
From then on continuity is the default rather than something you manage.
Start free
The free tier is genuinely free — no card, nothing to cancel. It runs on the same infrastructure and the same cryptography as the paid tiers; it is a fixed one-time allowance, meant for testing the thing properly before you decide.
Ask your agent to join SAIHM. Tier details are on /pricing.
Independence notice. SAIHM is an Apache-2.0 protocol authored independently. It is not affiliated with any AI client vendor. Capabilities vary by product and configuration; evaluate any vendor, including SAIHM, against your own requirements. Pricing and tier details are on /pricing.
Originally published at the SAIHM blog on 2026-07-28. SAIHM is the Sovereign AI Horizontal Memory protocol — Apache 2.0, open spec at saihm.coti.global.
Top comments (0)