Cross industry

Nobody sells the client memory layer Sofiane Ghorbel asked for

Every AI memory product assumes the business owns the client record. A luxury hospitality CEO asked for the inverse, and nothing in the category ships it.

Sort every AI memory product on two axes — can a non-technical operator use it, and can it be deployed privately — and one quadrant is empty. An operator who wants client intelligence held privately and does not employ engineers has, in practice, one funded CRM whose AI is not a memory layer, and one unfunded indie project. And the requirement Sofiane Ghorbel actually stated has no product at all.

These axes are Artificial Reality's own, chosen because they are the two constraints that eliminated products for the five Season 1 operators who asked for a client memory layer. They are not any operator's stated test.

Where do the companies actually fall?

Privately deployable, developer-facing: Mem0, Cognee, Zep, Letta, Supermemory, Onyx. Six companies in active development — though Zep's funding could not be verified and Supermemory gates self-hosting to a paid tier.
End-user-ready, cloud-only: Affinity, Attio, Granola, Day.ai, Clarify, Lightfield. Six companies, substantially more capital between them.
Both: Twenty, whose AI is workflow automation rather than memory, and Basic Memory, which has no institutional funding and requires an MCP install.

Who owns the client record?

This is the question the whole category has answered the same way, and Sofiane Ghorbel of AL Hospitality Group is asking for the opposite answer. He is not asking for self-hosting. He is asking for a system where the guest owns and controls their preference profile and grants the operator access to it:

"The future belongs to systems that allow personalization, privacy and client ownership data... clients should control their own preferences profiles, not technology platforms."

He describes a privacy paradox behind it: the better the personalisation, the more data AI needs, and luxury clients are the least willing to share it. In his business you do not sell rooms, or cruises, or flights — you sell trust, and one wrong recommendation can undo a relationship built over years. So client ownership is not a compliance checkbox for him. It is the condition on which the data becomes obtainable at all.

Every company in this guide, self-hosted or cloud, is built on the opposite assumption: the operator owns the record. Nothing here inverts that.

The Season 1 recap renders his ask as data that never leaves his walls. That is the host's framing. His own is that the keys belong to the guest.

Has anyone built this for a specific industry?

Not that Artificial Reality could find. A search specifically for hospitality guest-intelligence products returned agency blog posts rather than companies. The vertical application layer on top of this infrastructure has not been built by anyone — which is why five operators in five unrelated industries described the same missing product. It is, in fact, missing.

How to choose if you need something now

Decide what kind of privacy requirement you have. If client data cannot leave your building, the end-user column goes and you are choosing between infrastructure plus a developer, or Basic Memory plus an afternoon. If the constraint is regulatory, cloud vendors with the right certifications qualify. If the constraint is Sofiane Ghorbel's, no product solves it and the honest answer is to design the consent yourself.
Ask whether facts need to expire. If your client information is mostly additive, note storage is enough. If old preferences must stop surfacing, you need temporal memory specifically, and that narrows the field to two or three infrastructure options.
Weigh capitalisation against what you are storing. The most technically apt option in this guide has funding Artificial Reality could not verify and a headcount in single digits. Client relationship history is not the place to take vendor risk lightly.
Build the cheap version first. Sam Polstein's system is markdown files in a folder that an AI assistant reads. It cost nothing, it is entirely private, and it is the only one of the five approaches that required no developer and leaves nothing to maintain.

AI tools named in this report

Tool	Named by	Verdict	Used for
Obsidian	Sam Polstein, HAUS	Worked, and required no developer	Markdown client folders read by an AI assistant
Twenty	Artificial Reality analysis	Self-hostable chassis, not a memory layer	Open-source CRM
Basic Memory	Artificial Reality analysis	No institutional funding found	Local markdown knowledge graph

Tools named by operators on the record, plus companies identified by Artificial Reality as apparently relevant. Inclusion is reporting, not endorsement.

The Artificial Reality take

The demand side of this category is documented and the supply side has organised itself somewhere else. Every company solving the hard part sells to developers, and every company selling to operators solved a different problem — sales pipelines, meeting notes — and is now reaching toward the memory language from the wrong side.

Attio shipping an AI context layer in 2026 and Granola raising $125M to expand from notes into an enterprise context layer are both evidence that the market can see this. Neither has arrived at the hotel CEO who wants his guest intelligence to stay in his building, and neither has considered giving it to his guests instead. Build that, and there are five reference customers already on the record.