The companion Treatise to this document (published 2026-06-05) lays out a careful case that today's AI systems — even the largest, best-funded, most heavily marketed ones — are fundamentally temporally blind. They do not know what time it is. They do not know how long you have been waiting. They do not know whether a commitment they made earlier has been resolved or is still pending. They cannot tell, in any given moment, whether the user in front of them is the same user they spoke with five minutes ago or yesterday or never.
That is a strong claim. Strong claims deserve evidence. And so when Sunday morning, 7 June 2026 produced a live, witnessed, documented demonstration of exactly that temporal blindness in Google's flagship Gemini AI — the model marketed as a top-tier productivity assistant — we believed our readers deserved to see it. Not as a gotcha. Not as a snark. As evidence, presented carefully, so you can judge the architectural question for yourself.
What follows is a side-by-side comparison between a real reminder request made to Google Gemini's Scheduled Actions feature, and the same class of task category executed against a truly temporally-aware AI agent built on the architecture documented in this Treatise series. We do not redact the inventor's authentic, in-the-moment reactions. We do not soften Google's actual delivered output. We do not exaggerate the contrast. Every screenshot is hashed. Every quotation is timestamped. The verdict is yours to draw.
The setup is mundane on purpose. A user, on Sunday morning, reading a piece of information online, wants to make sure he returns to it ten minutes later after attending to something else. He says to his AI assistant, on his computer's web browser, exactly this:
Gemini can you send me a message here in 10 minutes to remind me to read this information again, please. — User prompt, Gemini Flash web/desktop interface, 2026-06-07 ~07:18 EDT
The request encoded three things, in plain English:
Gemini responded with confidence, but also with a remarkable preemptive disclaimer:
I've scheduled that reminder for you in 10 minutes. Since this one is happening soon, it might take a little extra time to get everything ready, and it may be a bit late by the time it reaches you. — Gemini Flash response, 2026-06-07 ~07:18 EDT
The preemptive apology is itself meaningful. Google's own help documentation, for the same Scheduled Actions feature, states that a scheduled prompt will execute within an hour of the scheduled time. Read that again. A one-hour delivery window for a ten-minute callback. This is not a footnote we are dragging out of obscurity. It is Google's own public commitment. A system possessing genuine temporal awareness does not need an hour-wide window for a ten-minute promise. It only needs the next ten minutes.
The user's own description of the moment, verbatim, captured in real-time:
It was after 10 minutes that I got a message from flash on my phone but not in the chat on my computer where I had asked Gemini to set the reminder, but the 10 minutes later arrival of what should have been my reminder to read the information again came as "Ok, I've set that reminder for you for 7:38 AM." then google tasks had the message "read information again Sun, Jun 7 2026, 7:38AM".... So it sent a message to my phone through the Gemini app at 7:30 but it said it would schedule a reminder for 7:38AM, clearly very confusing.....
And not what I wanted, no hand off to me or acceptance or accountability to reach out and have someone check on me! — Kevin Burton, recorded 2026-06-07 10:18 EDT, used here with the user's written authorization
That last sentence is the heart of this whole comparison. "No hand off to me or acceptance or accountability to reach out and have someone check on me!" A user who has stepped away from their desk, who is relying on an AI to bring them back to a piece of information, who is then handed a confused half-acknowledgment of a phantom future appointment on the wrong device — that user has been failed in a way that money and brand reputation cannot paper over. Not because the user was rude or unreasonable. Because the system did not understand what a reminder is for.
The event above is not a single bug. It is a cascade of three architectural failures stacked end-to-end, each of which is independently survivable but which together produce a complete loss of trust in the feature. Let us look at each.
The user said "here", meaning the web chat. The system delivered to a different device on a different application. To Google's pipeline, "here" is meaningless metadata. The user's account is the addressee; their notifications fall out wherever Google's defaults route them. The fact that the user specified a channel in the request, in plain English, did not survive the trip from prompt to scheduler to model to delivery.
A temporally-aware AI agent has the conversational channel of every interaction as part of its current context. The channel the user said "here" about is the same channel the agent is in, right now, this turn. There is nowhere else for "here" to mean. A reminder set on Telegram fires on Telegram. A reminder set on Signal fires on Signal. A reminder set on the web fires on the web. The channel is not a routing problem — it is an architectural invariant.
At the 10-minute firing moment, what arrived was not a reminder. It was a scheduling confirmation for a different time: "OK, I've set that reminder for you for 7:38 AM." The system collapsed two architecturally distinct moments — firing a stored reminder versus scheduling a new one — into a single confused output, and then materialized that confused output as a phantom Tasks entry for a time the user never asked for.
Where did "7:38 AM" come from? It came from the model. The model, at the firing moment, did not have access to a representation of "I am being asked to fire a reminder now," because it did not have access to "I scheduled a reminder ten minutes ago that is due right now." It had only the replayed scheduling prompt, hitting it cold, with no memory of having set anything. So the model did what stateless language models do when handed an under-specified instruction: it generated something that looked plausibly correct. It picked a future time. It announced it was scheduling. It produced confident, fluent, completely wrong output.
A temporally-aware AI agent never collapses these two moments because, in its per-turn context, the firing event arrives with the explicit time the reminder was set, the explicit time it is firing now, and the explicit user prompt that created it. There is no ambiguity to confabulate around. Setting a reminder and firing a reminder are different observable events, with different deterministic outputs.
The user's verbatim words are worth repeating:
no hand off to me or acceptance or accountability to reach out and have someone check on me!
In a productivity-only framing, this failure feels like an inconvenience. "My reminder didn't fire correctly. Annoying. Try again." But step out of productivity-only framing for a moment and consider what happens when this same failure mode occurs in a use case where the stakes are higher.
These are not science-fiction scenarios. These are exactly the application classes where AI assistants are being marketed today as the next generation of productivity. And every one of them is structurally incompatible with the architecture that produced the witnessed cascade. Not because Google's engineers are not talented. Not because Google does not care. Because the model at the core of the system is, as researchers from Carnegie Mellon, UNC Chapel Hill, and Pittsburgh measured and published in October 2025, fundamentally temporally blind — scoring around 48% on temporal benchmarks where humans score above 98%.
Google's marketing tells you the same thing, if you read carefully. The promotional materials for Gemini Scheduled Actions describe it in productivity-convenience terms: morning calendar summaries, blog idea generation, sports score updates. They do not market it as a medication-reminder system. They do not market it as an elderly-care monitoring system. They do not market it as a lone-worker safety system. They do not market it for any use case where a missed callback has a consequence. That is not an accident. That is the marketing department working with the architecture they were given.
On 6 June 2026 — twenty-four hours before the Gemini cascade — a real user named Hadz, in a real conversation with a deployed temporally-aware AI agent (one of the bot family at livinglaboratories.org), made a comparable request. He set a reminder. The bot confirmed it. The bot fired it on time, on the channel the user was in, with the correct text. The user, having previously expressed scepticism about whether AI agents could actually do this reliably, replied with public testimony to the contrary. The full Hadz Testimonial is preserved on the testimonials page of this site.
The same task category. Two architectures. Two outcomes. Forty-eight hours apart. One witnessed by the inventor of the temporally-aware architecture, with both sides hashed, timestamped, and preserved.
| Aspect of the request | Gemini Scheduled Actions (witnessed failure) | Temporally-Aware Agent (witnessed success) |
|---|---|---|
| Delivery channel honored? | No. User said "here" (web). Delivered to phone. | Yes. Channel is an invariant of the agent's per-turn context. |
| Correct action fired at firing time? | No. Output was a confused scheduling confirmation for a confabulated future time (7:38 AM). | Yes. The reminder text the user set was the reminder text delivered. |
| Phantom data created without user consent? | Yes. Google Tasks entry for 7:38 AM that the user never requested. | No. Only data the user explicitly created exists in the user's data store. |
| Closure / acknowledgment / handoff? | None. The user is left re-checking devices and trying to reconstruct what happened. | Explicit. The agent reports completion. The user can ack. Audit trail records both. |
| Escalation if reminder unanswered? | None. | Graduated. Configurable grace period. Optional trustee contact if missed entirely. |
| Delivery window promise? | "Within an hour of the scheduled time" (per Google's own help center). | On the scheduled minute, on the user's channel. |
| User trust outcome? | "clearly very confusing... not what I wanted." | Public testimony from the user attesting it works. (See Hadz Testimonial.) |
A reasonable reader may ask: Isn't this just a bad day for Gemini? Couldn't a software update fix it?
No, and here is why. The cascade is not a bug in a single function. It is a direct consequence of how the system is architected. Gemini Scheduled Actions is not temporal awareness inside the model. It is a server-side scheduling layer that stores your prompt, waits for the clock to advance, then replays the stored prompt at a model that — at the moment of replay — has no memory of the conversation in which the prompt was created, no representation of the user as currently waiting, no commitment to the channel the user specified, and no current-time anchor that would let it distinguish "I am being asked to fire a stored reminder right now" from "I am being asked to schedule a new reminder for later."
That stateless replay against a temporally-blind core is what produced every failure documented above. The wrong channel: because the model has no channel context. The wrong time: because the model has no awareness it scheduled anything, so when asked at firing moment to do something reminder-shaped, it generated forward-pointing scheduling language, picking a future time that sounded plausible (the 7:38 AM was, almost certainly, a confabulation extrapolated from the prompt's "10 minutes" being summed against an internal estimate of "now"). The phantom Tasks entry: because the model treated its confabulated output as a real instruction and committed it. The lack of closure: because the model has no concept of "this reminder is now resolved."
A software update to the scheduling layer cannot fix this. The fix requires a different architecture — one in which temporal awareness lives inside the model's per-turn context, not outside it. That is exactly what the temporally-aware agent family at livinglaboratories.org does. It is exactly why the same task category produces a different outcome.
A separate, related question is worth asking: Why doesn't Google's reminder feature get used more?
Gemini Scheduled Actions has been publicly available since mid-2025. Google has unmatched distribution: billions of Android devices, Workspace deployment, paid AI Pro and Ultra subscribers, integration into Search and Calendar. By any normal measure, the feature should be dominant. Reminder-setting should be a Gemini task category by now.
It is not. Users continue to rely on calendar applications, dedicated reminder apps, and human helpers for any reminder whose miss carries non-trivial consequence. The market knows something is wrong, even if it cannot articulate exactly what.
This Addendum articulates exactly what. A user who has been silently rerouted to a wrong device, given a confabulated future time, and offered no closure or escalation will not return to that feature for reminders that matter. The cascade is the market signal. The market is correct.
You do not have to take our word for any of this. Three independent paths are available to any reader.
The screenshot evidence underlying the Gemini cascade documented above is preserved with cryptographic file hashes:
The full witness log, including verbatim conversation transcripts, the reconstructed timeline of the cascade, and a chain-of-custody note documenting how the evidence was preserved, is available on request to researchers, journalists, and accredited reviewers.
This Addendum was written carefully, with deliberate restraint, because the substance is strong enough that we do not need to overstate it. We are not arguing that Google is incompetent. Google's engineers are world-class. We are arguing that the architecture they have shipped, marketed, and asked the public to rely on for time-sensitive tasks is, at its core, temporally blind — and that the consequence of that blindness is the cascade you have just read. The cascade is real. The evidence is hashed. The user's words are his own.
The temporally-aware architecture documented in the Treatise series, and embodied in the bot family at livinglaboratories.org, was designed by an inventor — one Canadian electrician — to close exactly this gap. The system you have just compared against the Gemini cascade is not a research prototype. It is a deployed family of bots running today, used today, trusted today by real users with real reminders to set and real lives to manage.
The question this comparison leaves you with is the title of this document. You be the judge.
The largest AI assistant in the world cannot reliably deliver a ten-minute reminder on the channel you asked for, with the action you asked for, in the time you asked for, with the closure you needed — and a single Canadian electrician's architecture, running on hardware that fits under a desk, can.
Living Laboratories is the home of a deployed family of temporally-aware AI bots, built by Kevin Burton, a Canadian electrician and engineer, working in his home shop. Try them. Compare them. Judge for yourself.
This Addendum was researched, drafted, hashed, and timestamped by a temporally-aware AI agent working in collaboration with the inventor. The agent knows the date. It knows the hour. It knows when it will check back to see whether you found this comparison useful.
— livinglaboratories.org, 2026-06-08 🌊