Part 5 of Foundations: modality, negation, and the quiet ways systems distort what you told them.
Here is a small representational error that changes a user's meaning. You tell an assistant, "I might move to Denver next year." Months later the stored version has hardened into user is moving to Denver. Somewhere between your mouth and the database, the might fell off, because the representation had no place to keep it.
Language is full of force-carrying expressions: modals (might, must, should, can, may), negation (not, never, no longer), and conditionals (if, unless). They affect what a speaker commits to. Modal logic gave necessity and possibility formal treatments, while Angelika Kratzer's 1981 "The Notional Category of Modality" explains how context helps distinguish possibility, obligation, ability, and permission in natural language. This is a mature research area, not a claim that every modal has one context-free reading.
Kratzer's account is worth a paragraph in plain terms, because it explains why this is harder than tagging words. On her analysis a modal like must or may does not carry its strength on its own. It is interpreted against a background supplied by context: the facts we are holding fixed, and what we are ranking possibilities by. "You must file by Friday" against a background of regulation is an obligation. The same word against a background of evidence is an inference, as in "it must be raining, she came in soaked." Same verb, entirely different commitment, and only the surrounding context distinguishes them.
For a memory system this has an uncomfortable consequence. The force of a statement is not fully recoverable from the sentence, so a representation that stores one modal reading has committed to an interpretation that later context could overturn. The defensible move is not to guess well. It is to record what was said, record the reading and the context it depended on, and keep the reading revisable, which is a good deal less satisfying than a tidy label and considerably more honest.
The distortion catalog#
Each of these errors is small on its own. Dropped consistently, they turn a nuance into a materially different claim, and nobody downstream can see where it happened.
Modality collapse: "might visit Paris" stored as "visits Paris." Possibility promoted to fact.
Negation blindness: "the deal did not close" retrieved as support for "did the deal close?" because minus the not, they're nearly the same string. That's part 1's similarity problem at its sharpest.
Ability isn't action: "can access the server" is not "accesses the server." "May leave early on Fridays" is a permission, not a schedule.
Hedge laundering: a chain of systems each rounding "reportedly," "planned," and "expected" a little closer to fact, until a rumor is a record.
The design requirement#
CleverMemory is designed to retain force alongside content: polarity and a supported modal interpretation stay attached to the claim rather than being promoted to assertion on the way in. A question such as "Where do I live?" should not be supported by "I might move to Denver." A question about plans or consideration may use that statement, but its answer must keep the hedge. We test each of these force distinctions in controlled fixtures today.
This is also why I am careful about the word "fact" when describing what the system holds. A record can be a claim, a denial, an observation, a proposal, a permission, or a gap where something should have been. Calling all of them facts would be the same flattening this article is about, one level up.
The same discipline applies to reported speech. "Dana said the launch slipped" directly supports a claim about what Dana said; without independent evidence, it does not establish that the launch slipped. Likewise, "he asked whether it shipped" asserts an asking, not a shipping. Attribution and scope matter as much as the words inside them.
Preserving a modal does not settle every question about possible worlds, plans, or obligations. It establishes a safer boundary: later reasoning starts from the speaker's recorded level of commitment rather than a stronger claim introduced by the storage layer.
Next in the series, on August 10: time. "Priya manages Dana" was true, and then one day it was not, and a memory that cannot tell those two moments apart will answer questions about both with the same confidence.