Introducing Universal Context

Every restaurant and event venue we work with runs on the same accidental architecture. There is an inbox full of inquiries, a second mailbox the events team checks on Fridays, a calendar of tastings and site visits, a reservation system, a spreadsheet of event leads, and a folder of menus and price lists. Each one works. None of them know the others exist.

The result is a strange kind of amnesia. A guest can email the events address to ask about the private dining room, write to the general inbox with a follow-up question the next morning, come in for a tasting that afternoon, and send their menu choice from a personal address that evening — and to the business's software, that is four strangers. The person answering the events inbox has no idea the other thread exists. The reply from the general inbox doesn't know a tasting happened. Nobody connects the first inquiry to the deposit that eventually lands.

For a hospitality business, that amnesia is expensive in a very specific way: the entire product is remembering people. A venue that remembers you is hospitality; a venue that asks you to repeat your date, headcount, and dietary notes for the third time is a call center.

We built Winston around a single opinion: before an AI agent can usefully run a front office, every interaction the business has ever had with a guest must live in one shared, structured layer that the agent reads before it says a word and writes to after every action. We call that layer universal context. Today I want to explain how it works, why we built it into the foundation instead of bolting it on, and what it makes possible.

The front office is a pile of silos

The obvious way to build an AI product for hospitality is channel by channel: an email assistant here, a phone bot there, an ads tool on the side. It's obvious because it matches how the tools are bought. It's wrong because it matches how the tools are bought.

Channel-shaped AI inherits channel-shaped amnesia. An email assistant that only sees one inbox will cheerfully quote a menu to someone who was told a different price from the events mailbox yesterday. A booking bot that only sees the reservation system will ask a bride for details she already typed into her inquiry email. Each assistant is locally competent and globally embarrassing — and in hospitality, globally embarrassing is fatal, because the guest experiences the business as one entity even when the software is twelve.

There is also a quieter cost. Every unlinked interaction is training data the business throws away. Which kinds of inquiry produce guests who actually show up? Which inquiry sources convert to deposits? If the inquiry, the tasting, and the booking live in three systems, those questions are unanswerable — not hard, unanswerable.

So we inverted the build order. Winston — our agent — was never a set of channel bots. It is one agent with one memory, and the channels are just the places it listens and speaks.

One timeline per guest

Universal context is, concretely, three things.

First, ingestion. Every source Winston touches — the AI inbox across Gmail and Outlook, Google and Outlook calendars, the event pipeline, email sequences, the documents the business uploads — emits events into one stream. A calendar meeting produces its attendees, its time, and the record it belongs to. An email thread produces its messages and extracted intent. A sequence step produces its send and any reply. We normalize all of it into a single event shape: who, what channel, what happened, what it implies.

Second, identity resolution. Events arrive keyed on fragments — an email address, a company domain, a phone number in a signature, a calendar invite. The context layer's job is deciding that anna.k@… on last Tuesday's calendar invite and the Anna signing this morning's email from a new address are the same Anna. We resolve conservatively: exact identifiers merge automatically; probable matches are linked with a confidence score and surfaced for one-click confirmation rather than silently merged. A wrong merge is worse than a missed one — greeting a guest with someone else's history is the exact failure we exist to prevent.

Third, structured recall. A raw event log is not context; it's a haystack. On top of the timeline we maintain a compact, continuously updated summary per guest — preferences, open threads, commitments, upcoming dates — plus the business's own documents in Resources: menus, rooms, policies, pricing. When Winston engages, it doesn't embed-and-hope across everything ever recorded. It loads the guest's summary, the open loops, and only then searches the deep timeline if the conversation demands it. Fast path first, haystack second.

A word on boundaries, because a shared memory raises a fair question: who gets to read it? Universal context is scoped to the business, full stop. One venue's timeline never informs another venue's agent, never trains a shared model, and never leaves the tenant boundary. Inside the business, the same permissions that govern staff govern Winston: read-only connections stay read-only, and anything that sends or books runs at the autonomy level the operator chose — off, draft, or auto — per inbox. We treat that as an architectural property of the context layer, not a settings page, because trust in the memory is what makes everything downstream of it usable.

The other load-bearing decision is that this layer sits below every feature, not beside them. Nothing in Winston talks to a channel directly; everything reads from and writes to universal context. That constraint hurt early — the inbox would have shipped weeks sooner as a standalone — but it means every new channel we add is born with the memory of all the others, instead of starting a new silo we'd someday have to reconcile.

What Winston does with universal context

The payoff shows up as small moments that compound.

Every reply knows the whole history. When a guest writes to a different inbox about “the quote you sent,” Winston is already holding that quote — the date, the headcount, the menu tier — and can check availability and confirm the booking in that same reply, instead of promising to get back to them.

The inbox knows the last meeting. When the same guest emails two days after their tasting, the draft Winston proposes knows the tasting happened, because the calendar event landed on their record the moment it synced. Whether that draft sends automatically or waits for approval is the operator's setting, not ours — Winston drafts; the business decides its own autonomy level.

Reporting knows what a booking is. Because the inquiry, the quote, and the deposit sit on one timeline, reporting counts guests who actually booked, not inquiries. Forecasting in the event pipeline improves the same way: it reads real conversion behavior, not a hand-maintained stage field.

And the write-back loop is what makes it durable. Every action Winston takes — a reply sent, a reservation created, a record updated — becomes a new event in the same timeline. The agent's actions and the guest's actions are one history. There is no separate “AI log” drifting away from reality.

None of these features required new intelligence. They required the model to stop being the bottleneck and the context to stop being the bottleneck instead — and then removing it. Our experience matches what the whole industry keeps rediscovering: past a threshold, agent quality is context quality.

Where universal context goes next

Two directions from here.

Proactive context. Today Winston reads context when a guest reaches out. Increasingly it will act on context when nothing happens: the couple whose tasting was three weeks ago and whose contract is unsigned; the regular who hasn't booked since spring. Silence is an event too, and it belongs on the timeline.

Context as the product's edge. Owners can already ask Winston questions across their whole business — “what did we quote the Lindqvist party?”, “which campaigns filled last month's Fridays?” — and get answers grounded in the same timeline the agent works from, not a separate analytics copy.

The models will keep improving underneath us, and that's precisely the point of this architecture: every model upgrade lands on top of complete memory. A smarter agent with amnesia is still a stranger. Universal context is how Winston stops being software the business uses and starts being staff the business trusts — the one member of the team who was present for every conversation, remembers all of it, and never leaves.

Winston is live today across email, calendars, and events for restaurants, venues, and caterers. If you'd like to see your own front office with one memory, book a demo.