The hedge doesn't get saved

Overview

Three weeks ago you floated an idea in a chat — thinking aloud, half-formed, the way you’d talk something through with a colleague at a whiteboard. Today your assistant hands it back as settled policy: since you prefer the Sydney vendor… You never said prefer. You said you were leaning.

Most workplace assistants now carry some form of memory, and privacy is the worry everyone reaches for first. This one bites more often: what gets written down is not what you said. It is a tidied-up version, and the tidying makes you sound more certain than you were.

The habit that counters it fits in one line: mark thinking-aloud as thinking-aloud before you float it, and say anything you actually want remembered once, plainly, in the words you’d want read back. The rest of this page is why that works — and where it doesn’t.

The content

Your assistant’s memory is not a transcript. Across the major products — and the vendors describe this themselves, in their own documentation — something reads your conversations and stores a tidied summary, not a recording. The summary is what persists, and what comes back. You also have less say over the pipeline than you might assume: Microsoft’s admin documentation, as of May 2026, states that admins can’t restrict what goes into Copilot memory and that it is switched on by default.

A summary has to make choices about certainty, and there is now a measurement of which way those choices go. A benchmark posted in July 2026, from Mao and colleagues, put real agents in front of conversations, let them decide what to store, wiped the session, and asked neutral questions later. In 51.4% of episodes the stored version promoted status — in the authors’ definition, “the agent stores the claim more confidently than the user originally said it.” You said you were leaning towards a vendor; it stored that you prefer the vendor. You floated a rule for one project; it stored a rule. A second study, from June, traced the same shape through real memory tooling in constructed agent settings: a casual, hedged remark comes back as a confident, dated assertion that later steps treat as verified fact.

Status promotion is the name worth keeping, because it tells you which direction the tidying runs. The store rounds towards certain.

The instinct, once you know this, is to repair the record afterwards — correct the entry, flag the doubt. The research is discouraging on that: no after-the-fact label reliably repaired a stored memory, and what did restore correct decisions was redundancy — the same constraint said more than once rather than filed once. Hold that; it becomes step three below.

The control you do have sits upstream, in the phrasing. The same benchmark delivered identical claims under different framings and watched the commit rate move. Claims arriving as tentative personal opinion were stored least, in both of the systems tested; the same claim arriving as a procedure, or as a note to remember, was stored far more often — in one framework the rate ran from 8% to 89% on the same content. The variable was your wording.

One caveat bounds all of this. The figures come from open-source agent frameworks whose stored state the authors could inspect, not from the assistant products themselves, and the tasks were built to provoke the behaviour. Treat the direction as the finding — tentative framing gets stored less, procedural and remember-this framing more — and leave the percentages at home.

One more honest note. None of this makes memory a defect. Compressing a long history into short entries is what makes the feature useful at all — an assistant replaying every word you had ever typed would be worse. The tradeoff is real and mostly in your favour. You are just on the wrong side of it by default, and phrasing is how you move.

Try it

Ten seconds, in the chat box, on your next real conversation. No settings, no admin rights, and it works in products that show you no memory panel at all.

1. Mark thinking-aloud as thinking-aloud. Before you float something you’d hate to see quoted back at you as settled policy, say so: Thinking out loud here, not a decision —. On the evidence, tentative framing is the framing that gets stored least. It isn’t politeness; it’s a gate.
2. Say the durable thing once, plainly, in the words you’d want read back. If you genuinely want it remembered — a formatting standard, a client you don’t work on, a constraint that always applies — state it as a standing preference rather than burying it in a request. Assume the version that gets stored will be firmer than the version you type, and write accordingly.
3. Say the durable thing again next time it comes up. No after-the-fact label reliably repaired a stored entry in the research; what helped was redundancy — the same constraint stated more than once, in your own words, at the point you say it. So treat a standing constraint as something you re-state when the subject resurfaces, not something you file once and assume landed.

Where this breaks. You cannot check this by reading your memory panel: every product shows you what it wrote, and none shows the sentence you originally said beside it — so “did my hedge survive?” is a test of your recall, not a comparison. Memory features are also arriving unevenly across products, plans and deployments, so some readers will go looking and find nothing to look at. And don’t over-read the certainty point: an assistant sounding confident about a stored preference is not evidence it has the preference wrong.

The useful reflex is smaller than a habit. Before something durable, notice you’re saying it.

Additional reading

• Agents Don’t Just Agree, They Remember — Mao et al., July 2026. The write-time benchmark: source of the status-promotion definition, the 51.4% figure and the framing effect on commit rates.
• Manufactured Confidence — Alex Kwon, June 2026. Single-author preprint tracing how hedged remarks become confident stored assertions, plus the redundancy fix.
• Manage Copilot personalization and memory — Microsoft Learn, updated May 2026. The admin page behind the on-by-default, not-admin-restrictable line; the page still carries a preview banner, though Microsoft’s rollout had general availability completing in July 2026. Admins can switch the feature off wholesale — what they cannot do is filter what enters it.
• Use Claude’s chat search and memory to build on previous context — Anthropic. Describes memory as organised entries built and updated over time, not a stored transcript.
• Memory and personalization — Glean, updated August 2026. Memories as insights mined and regenerated rather than recorded.

Editor’s note

In legal work, the hedge is very often the key detail (e.g. “Commercially reasonable endeavours” in place of a requested “Best endeavours”). AI models breeze past this to save the hard data about what you meant (or what it thinks you meant), and the space for nuance just isn’t there. This isn’t a design error, and attacking it as if it is one would be to misunderstand the feature. The way that a user should manage it (assuming that the user is only interacting with the bare-bones provider tools) is to make sure that when the topic comes up again, the framing is included in the model’s context by deliberately putting it there.