What an AI memory layer is, and why five operators asked for one

An AI memory layer is a system that stores structured, persistent facts about a person or engagement, updates them as new information arrives, retires facts that are no longer true, and makes the current state readable by an AI assistant at the moment of use. Five operators in five unrelated industries described it to Artificial Reality across Season 1, and none knew the others had asked.

A yacht charter broker wants a system that remembers which wine a guest drank and when the children's birthdays fall. A luxury hotel CEO wants scattered guest intelligence made usable. An M&A broker wants a deal room per client. Galina Fendikevich, who hosts Artificial Reality, the AI buyer-evidence podcast, put them together in the Season 1 recap:

"Those are all the same product. A private, structured, but dynamic memory layer for client profiles in high-trust businesses. Way beyond a CRM. And nobody's built the clean version of this."

A note on the count: the recap episode says six executives. The list it reads names five, and this cluster follows the five named.

How is a memory layer different from a CRM?

A CRM models a pipeline — deals, stages, opportunities. A memory layer models a relationship over time: structured facts about a person that update as things change and retire when they stop being true. It differs from note storage in that it maintains structure and currency rather than accumulating text. The operators who asked run businesses with clients, not pipelines — which is why the recap reached for "way beyond a CRM."

Who asked, and what each one asked for

Jennifer Kerum, yacht charter (Anchor Croatia). Repeat guests, multi-year relationships, and the details that matter enormously and live nowhere: the wine, the birthdays, what went wrong last time. Her broader complaint is that charter companies and yachting "almost get forgotten in the AI world."

Sofiane Ghorbel, luxury hospitality (AL Hospitality Group). The whole guest journey as the unit — "start from the inquiry, discovery, curation, booking, travel, follow up, loyalty, repeat business" — with follow-up, once a client is travelling, as the stage where the tool is missing. "I don't have this tool till now. I hope that we will find it, but it will be great to have it."

Sam Polstein, deep tech communications (HAUS). The only one of the five who solved it, by hand. "I think about context engineering a lot more than I think about prompt engineering." Client history, voice and examples of good writing sit as markdown in Obsidian, readable the second he points an AI assistant at a client folder. His complaint about existing contact tools is currency: a three-week data lag makes them useless for this work.

Christine McDannell, M&A (The Magnolia Firm). A deal room per client, holding the full context of a transaction, instead of assembling it by hand across tools. Asked whether she wanted real M&A deal infrastructure rather than duct-taping her own, she said it would be rad.

Dr. Nikki Siso, health coaching. Built a multi-skill client assessment system in Claude with no technical background, then hired a developer to wire the back end together, because the intake brain she needed did not exist as a product.

What would the product actually have to do?

Six requirements come straight from the operators:

1. Hold per-person facts that persist across engagements — the wine, the birthday, the thing that went wrong — and surface them before the next interaction rather than on request.
2. Ingest what already exists in unstructured form. Handwritten notes, email threads, past documents. Sofiane Ghorbel's guest intelligence exists; it is scattered and unusable.
3. Make privacy the condition of getting the data at all. His stated requirement is client-owned preference profiles: the guest controls what the operator sees.
4. Stay current. Sam Polstein does not care what a reporter wrote six months ago. He needs what they will write next.
5. Be readable by an AI assistant. The memory is only useful if the model can consume it.
6. Be structured per client, not per deal. A deal room and an intake brain are both containers organised around a person, not a pipeline stage.

Artificial Reality adds four more, and the first is the hard one. Facts must be able to expire. A guest who drank Chablis in 2023 and switched to Burgundy in 2025 has one current preference and one historical one, and a memory layer that only accumulates becomes wrong over time. Then: usable without a developer, because all five run businesses and three built their own version or tried to. Exportable, because the accumulated client record is the asset in each of these businesses and cannot be hostage to a vendor. And auditable access, because in hospitality and M&A specifically, who looked at a client record is itself a compliance question.

AI tools named in this report

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

What a builder should take from this

Five operators, five industries, no contact with one another, one product. That is about as clean a market signal as customer discovery produces, and the specification is unusually complete: temporal per-person memory, ingesting unstructured history, readable by an assistant, deployable privately, sold to someone who has never opened a terminal.

So the question for a founder is not whether the demand is real. It is which half of the problem you already solve, and whether you have been ignoring the other half.

Where this comes from

S1E8: Why 99% of AI products fail — the full interview with Galina Fendikevich. Listen or watch: YouTube, Spotify or Apple Podcasts.