Your Notes Say \"Tomorrow\" — the AI Has No Idea When That Was

A note that reads "ship this tomorrow" is exact to you and empty to a language model. The model reading it has no clock, and it cannot see the day you typed the word. So "tomorrow" resolves to nothing. Write the absolute date instead, 2026-07-22, and the sentence still means something the day a machine reads it back.

One correction before the argument, because the title overreaches. A model can resolve "tomorrow" — if something hands it today's date and the note carries the day it was written. The chat apps do the first half for you. Anthropic's own documentation says Claude's web interface "and mobile apps use a system prompt to provide up-to-date information, such as the current date, to Claude at the start of every conversation."¹ The interface fills the blank because the model has none to fill.

The failure this post is about is narrower and far more common: a note read on its own, later, with no date supplied and none inside it. Then "tomorrow" points at a day nobody recorded.

What most people believe: plain text explains itself

The reasonable belief is that a plain-text note explains itself. You wrote it in your own words, the words are right there, and any reader can pick it up and understand it. For a human reader that mostly holds — because the human brings a second source the note does not contain: the memory of when they wrote it.

Read "call her back tomorrow" three weeks later and you still know what it meant, because you remember the Tuesday you typed it. The note is half the message. Your memory is the other half. Relative time words — tomorrow, last Friday, next week, in two days, by end of day — are pointers into that second half, the part that never made it onto the page.

The note points. You supply the origin.

Why it fails: the reader has no clock and was not there

A language model brings no second half. It has no wall clock, and it cannot see the file's authoring date, which lives outside the text, in metadata the model was never handed. Give it "call her back tomorrow" with no other anchor, and the most it can do is guess which day "tomorrow" was counted from.

The deeper reason it cannot simply reason its way there is that a model holds no number line of time. It holds statistical associations between tokens. Tan, Ng and Bing, in TempReason (ACL 2023), state the mechanism plainly: "language models only learn the co-occurrences of time expressions and their context."² The same paper shows how literally a model splits a date: in one tokenizer "the year '2021' is tokenized as '20' and '21',"³ two fragments rather than a point on a timeline.

There is no calendar inside the model. There is only text that tends to sit near other text.

A relative expression is therefore the worst case. It names no absolute point at all — only an offset from a moment the model cannot see. "Tomorrow" is a subtraction with the first number missing.

What the research shows: relative time is where models slip

Even when you hand a model a real anchor, temporal reasoning is where it slips. Four benchmarks from four research groups point one direction: the finer the granularity and the more an answer leans on relative reasoning, the worse the model does. The weakness is documented, but documented weakness is not the claim that a model can never read a date.

Look at how the accuracy collapses as the question tightens. TempReason reports that "ChatGPT is able to achieve 99.6 EM on year prediction, whereas it can only achieve 30.5 EM on month prediction."⁴ Right about the year almost every time; right about the month barely a third of the time.

The same shape shows up in TimeQA, a NeurIPS 2021 dataset for time-sensitive questions: "The best-performing model FiD can only achieve 46% accuracy, still far behind the human performance of 87%."⁵ Its authors are blunt about the cause: "these models are still lacking the ability to perform consistent temporal reasoning."⁶

You could object that benchmarks only prove models memorized the wrong web pages. A Google DeepMind and Research team built Test of Time to close exactly that gap, using synthetic data because prior benchmarks "often rely on real-world data that LLMs may have encountered during pre-training."⁷ Strip out the memorization and the fragility survives: "they remain susceptible to errors, particularly in temporal reasoning tasks involving complex temporal logic."⁸

The reasoning breaks down as reasoning, not as recall.

Here is the finding that is easy to misread, so read it slowly. In Temporal Referential Consistency (EMNLP 2025), Bajpai and Chakraborty report that "LLMs demonstrate superior performance in handling chronological temporal references as opposed to absolute temporal references, thereby resulting in greater temporal referential inconsistency across languages."⁹ Their "chronological reference" is an event anchor the model already knows — "before he joined the company," "during the seminary years." That is the opposite of a note's bare "tomorrow," which hangs on a day the model was never present for.

The lesson is not that relative words are easy. It is that the form of the reference changes the answer: the same date, written differently, moves the model's accuracy. A note's "tomorrow" is the hardest form there is, an offset from an origin no one wrote down. The categories the field uses make the gap concrete; the harder levels of reasoning, "time-event (L2), and event-event (L3),"¹⁰ all assume an anchor a bare note does not carry.

None of this says a model can never resolve a date. Supply today, keep the note's own date in the file, and "in two days" becomes arithmetic even a weak model often gets right. The durable problem is the note that travels without its date and gets read when you are not in the room to supply it.

What to do instead: write the date the standard way

The fix is not a smarter model. It is a note that carries its own anchor. Write the absolute date in the one format built to be unambiguous, ISO 8601, YYYY-MM-DD. So 2026-07-22, not "tomorrow," not "7/22," not "next Friday." A note that names its own dates can be read alone, by you or a machine, years later, and still resolve.

ISO 8601 exists for precisely this. The International Organization for Standardization describes it as "an internationally agreed way to represent dates: YYYY-MM-DD."¹¹ It removes the ambiguity that trips even humans: is 03/04 the fourth of March or the third of April? On the internet the same standard has a profile, RFC 3339, which Klyne and Newman define as "a profile of the ISO 8601 standard" giving an "unambiguous representation of some instant in time."¹² Every timestamp your tools already write is a version of this. The convention is older than the problem.

Five changes, each a keystroke more than the shortcut you were about to type:

1. Replace the relative word with the date. "Follow up tomorrow" becomes "Follow up 2026-07-23." The word "tomorrow" only ever worked next to a day you did not record.
2. Date the note itself. A single date: 2026-07-22 line at the top gives every "tomorrow" below it an origin to count from — and that date belongs in the file's text, not in fragile filesystem metadata, as argued in your note dates are filesystem metadata.
3. Prefer YYYY-MM-DD everywhere. It sorts in the right order as plain text, reads the same in every country, and never asks the reader to guess whether the month or the day comes first.
4. Make the date a field for tasks. due: 2026-07-30 beats "due end of month," which quietly means a different day every time it is read. The structured form is covered in due dates in a plain text task file.
5. Keep the two problems separate. Dating the language inside a note is a different job from deciding which of two contradictory notes is current — that second job, the supersede pointer, is the companion piece: date your notes so the AI knows which one is current.

Frequently Asked Questions

The questions below are the ones a note-owner arrives with after an assistant confidently misread a deadline that was perfectly clear to them. The short version: the model was not there when you wrote the note, cannot see when you wrote it, and reads "tomorrow" as a word with no day attached.

Why does AI get the dates and deadlines wrong when it reads my notes?

Because your notes use relative time — tomorrow, last Friday, next week — and the model reading them has no clock and cannot see the day you wrote them. TempReason names the root cause: models "only learn the co-occurrences of time expressions and their context,"² not a real timeline. Written absolute, 2026-07-22, the deadline stops depending on a day the model never saw.

Does the AI know when I wrote my note?

Not from the note alone. A file's authoring date lives in metadata outside the text, and the model is handed the text. Unless the date is written inside the note, or supplied in the prompt, the model has no way to know when "yesterday" was. This is why the durable place for the date is the note's own body, not its file properties.

Why doesn't ChatGPT know today's date?

Because a language model has no internal calendar; it predicts text, it does not run a clock. Chat apps work around this by injecting the date; Anthropic documents that its interface adds "the current date, to Claude at the start of every conversation."¹ Through the raw API, with nothing supplied, the model genuinely does not know what day it is.

Why does ChatGPT get the current date wrong?

Same mechanism. When the date is not supplied, the model guesses from patterns in its training text, and time literals are handled poorly — TempReason showed a tokenizer splitting "2021" into "20" and "21,"³ two fragments rather than a point in time. It is architecture, not a bug in one product; supply the date and the guesswork disappears.

What date format should I use in my notes?

ISO 8601: YYYY-MM-DD, so 2026-07-22. The standards body calls it "an internationally agreed way to represent dates,"¹¹ and its internet profile, RFC 3339, exists to give "an unambiguous representation of some instant in time."¹² It sorts correctly as text and reads the same everywhere; no reader, human or machine, has to guess the order.

Can AI understand "tomorrow" in my notes?

Only if it is given the origin to count from. Hand the model today's date and a note that carries its own date, and "tomorrow" becomes simple arithmetic it usually gets right. Read the note on its own, with neither, and "tomorrow" resolves to nothing — the offset is real but the starting point is missing.

A relative date is a promise that someone will remember the day it was made. You are not always that someone, and a machine never is. "Tomorrow" is a message to a reader who was standing exactly where you stood; write the date, and the note stops needing a witness. A note that names its own dates is one you can still resolve in five years — by a memory you no longer have, or a model that was never there.

Notes in MNMNOTE are plain Markdown you own, kept on your own device, where writing 2026-07-22 instead of "tomorrow" is a keystroke that outlives the day you typed it — mnmnote.com.