TL;DR
When an assistant "knows" more about you (ChatGPT memory, a profile, past chats), its answers don't automatically get better. The researchers compared model answers with and without your memory. Users picked the memory version about as often as the version without it. And the more insistent the "use personal context" instruction, the more often the memory-free answer won.
You know the pain: you once asked for explanations "in the language of your profession," and now you want a general overview of a topic. The model still tailors the answer to you, narrows it, and fills in things you never asked about. The reason is simple. The model sees facts about you in the context and treats them as an invitation to use them. It's bad at judging whether using them is appropriate. Automated judges (LLM judges and off-the-shelf reward models that score answers) err in the same direction: they almost always praise the memory answer, while people don't.
Practical takeaway: don't ask the model to "personalize as much as possible." Set a strict criterion ("use memory only if it noticeably improves the answer to the current question") and check the result yourself instead of handing the evaluation to an LLM judge. The authors also trained a small preference-prediction model. That's infrastructure you don't need as a reader, but its features hint at what to look for when you check.
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Method Blueprint
The paper has no single prompt method. It has three practical takeaways and an evaluation design you can reproduce by hand.
TAKEAWAY 1: Memory instruction → conservative ("only if it noticeably improves"), not aggressive
TAKEAWAY 2: In doubt → compare the answer WITH memory and WITHOUT it (same question), pick yourself
TAKEAWAY 3: Evaluation → don't trust an LLM judge on personalization; look at three signals:
• relevance of memory to the query
• proportionality of assumptions
• unfounded guesses about the user
All of this works in a regular chat, in a system prompt, or in an agent's instruction file. The steps fit into one or two requests.
Example
Task: A marketer runs a YouTube channel about selling on Amazon. That's stored in ChatGPT's memory. Now they're preparing a lecture for students and ask for a general breakdown of subscription models. They don't want half the answer to be about marketplaces and product listings.
Prompt (in custom instructions or at the start of a chat):
Rules for using memory about the user:
— Use information about me only if it noticeably improves the answer
to the current question specifically.
— If the question is general, answer generally, without tying it to my work or interests.
— Don't make assumptions about me that aren't in the question or in memory.
— If you're unsure whether to use memory, don't.
You can offer, in one line at the end, to adapt it to my context.
Question: Break down how subscription models work in business:
types, economics, common mistakes. I need a general overview for a lecture.
Result: The model gives a universal overview of subscription models with no "now imagine you're selling on Amazon." At the end there may be one line offering to adapt the material to your context. If the answer is still skewed, ask for a version "ignoring my memory" and compare the two texts.
Why It Works
The weakness. If facts about the user are sitting in the context, the model tends to use them because they're "right there." A "personalize more" instruction only amplifies the skew. In the study, the most aggressive mode produced the highest rate of people choosing the memory-free answer. Even the gentle instruction beat "no memory" only by a hair. A single general request to be selective isn't enough, so you have to sharpen it with criteria.
What the model can do. It follows explicit conditions well: "use it only if…", "answer general questions generally", "don't fill in the blanks." A conservative wording turns a fuzzy "be relevant" into a check the model can actually perform.
Control knobs: - Memory inclusion criterion: "only if it noticeably improves" → stricter, less intrusive. - Rule for general questions ("answer generally") → protects against narrowing the answer. - Ban on guesses about the user → fewer unfounded assumptions. - A one-line offer at the end instead of baked-in personalization → you decide whether you need the adaptation. - Request a "no-memory version" → a quick check of whether the answer got skewed.
One more finding. People preferred answers with more headings and greater length. Form influenced the choice more than anything else. A good-looking answer is easy to mistake for good use of memory, so compare answers of the same length and structure.
Prompt Template
Rules for using memory and personal context:
1. Use what you know about me ONLY if it noticeably
improves the answer to the CURRENT question.
2. If the question is general or unrelated to my context, answer generally,
without tying it to me.
3. Don't make assumptions about me that aren't in the question or in memory.
4. In doubt, don't use it. You can offer, in one line at the end,
to adapt it to my context.
{question}
Checking the answer (for you, not for an LLM judge):
Here's the answer to my question: {answer}.
List separately:
a) which facts about me it uses;
b) for each fact, whether this question needs it (yes/no and why);
c) which assumptions about me in the answer aren't backed by my question or memory.
Then give a version of the answer without those facts and assumptions.
Fill in {question} and {answer}. The second prompt is built from the paper's three signals: memory relevance, proportionality of assumptions, and unfounded guesses. The call on "which version is better" stays with you. In the study, an LLM judge with a rubric barely improved accuracy, so on its own it's no guarantee.
Limitations
⚠️ The effect is small: Even the conservative instruction gave only a slight edge over the memory-free answer. It's not a magic setting, just a way to avoid making the answer worse.
⚠️ Different models behave differently: For GPT and Claude, memory won more often; for Gemini, it lost. Test on your own model.
⚠️ The pairs were hand-picked: Participants were shown queries where the answers with and without memory differed noticeably. On a normal stream of queries the gap may be smaller or different.
⚠️ Human preferences are noisy: Three days later, the same people changed their choice in about a third of cases. For weak preferences, agreement was even lower. That's why only strong preferences were used to evaluate the models.
⚠️ You don't get the trained model: The headline result (predicting preferences with a small trained model) requires labeled data and code. For a practitioner it's not a method, just a hint about which signals matter.
⚠️ Don't expect objectivity from an LLM judge: It inflates memory answers and looks little like real human preferences. Don't delegate personalization quality assessment to it.
⚠️ Form effects: Headings and length strongly sway the choice. In real work it's hard to separate "memory was used well" from "the answer is just longer and more structured."
The available text of the paper cuts off at the breakdown of features into blocks, so the details of the final experiments aren't covered.
How They Tested It
The researchers recruited 87 participants on the Prolific platform. Each ran a special prompt in their own ChatGPT that collected a profile, topic preferences, and about a hundred recent queries. For each query in that set, nine relevant memory entries were pulled in. Then one of three current models (GPT-5.2, Gemini-3.1-Pro, Claude-Opus-5) generated two answers: with memory and without. Memory was injected in three modes: conservative, standard, and aggressive. Each participant saw a pair of answers to their own question using their own memory and chose on a five-point scale which one they'd rather get.
They collected about a thousand ratings; after cleaning, 821 remained from 87 people. The result was unexpected. With the standard instruction, memory was chosen in 38.7% of cases and no-memory in 41.9%. In aggressive mode, the memory-free answer won 49.3% of the time. The conclusions are sensible: "the more personal, the better" is a myth, and a "personalize" instruction acts as a skew amplifier.
Then they tested automated judges. Plain GPT-5.2 guessed the human's choice 58% of the time, a judge with an off-the-shelf rubric 59.6%, and the baseline strategy "always pick no-memory" scored 55%. Judges systematically picked the memory answer more often than people did. A trained model on 12 features (logistic regression on top of LLM rubric scores) reached 72.8%, nearly matching the people's own agreement on re-rating (73.9%). The strongest features: number of headings, relative length, memory relevance, proportionality of assumptions. Unfounded assumptions about the user had a negative weight.
For practice, this is what matters: the key quality criterion is appropriateness, and the model can't judge it on its own. You need a human in the loop or clear rules.
Adaptations and Extrapolations
🔧 Technique: two answers in one chat → a quick skew check
Replicates the study design, by hand:
Answer my question twice.
Version A: use everything you know about me.
Version B: answer the way you'd answer a stranger.
Both versions should be roughly the same length and structure.
Question: {question}
Equal length matters because in the study, form strongly swayed the choice. Pick the version yourself. If you prefer B more often, memory in your instructions is getting in the way.
Extrapolation (not from the paper): rules for the instruction file of an agent with memory (CLAUDE.md, AGENTS.md, a custom GPT's system prompt).
## Using context about the user
- Apply information about the user only if it noticeably changes
the correct answer to the current task.
- For general questions, give a general answer.
- Don't infer assumptions from memory that the user didn't ask for.
- If you applied a personal fact, name it in one line so
the user can turn it off.
I added the last point as a way to check the output. The study didn't test it.
Resources
- Memory-Aware Reward Modeling: Learning Human Preferences for LLM Memory Use
- Authors: Yijiang River Dong, Zexue He, Cheng Yu (equal contribution), Tiancheng Hu, Michael J Ryan, Jan Batzner, Zheng Hui, Shaolong Wu, Orestis Papakyriakopoulos, Alex Pentland, Nigel Collier, Jiaxin Pei
- Universities: Cambridge, Stanford, TU Munich, Harvard, UT Austin
- Related work cited in the paper: OP-Bench (over-personalization), PersistBench (memory leaking across topics), LongMemEval, LoCoMo, PersonaMem