TL;DR
When an LLM is given a constraint ("answer as if it's 2015", "you're a junior with no experience", "write only about what you know for sure"), it often breaks it without anyone noticing. Especially when the link between the request and the forbidden knowledge is indirect: not a direct question about an event, but an inference from it. Two steps fix this: Self-Recall (SR) and Question-Recall (QR). The model first states the constraint in its own voice, then recalls relevant facts within that constraint.
The key finding is a surprise: the standard "let's think step by step" (Chain-of-Thought) hurts constraint adherence. The reason: step-by-step reasoning prompts the model to recall facts that are formally off-limits. The model "thinks out loud", involuntarily drags forbidden knowledge into the reasoning chain, and then inevitably uses it in the answer.
SR takes a different route: instead of reasoning, a short self-declaration of the constraint. One sentence in which the model states the rule in its own voice. QR adds a second step: one sentence with relevant facts and their years, laying down the "allowed" knowledge before the answer. Together: SR → QR. Two sentences added to the prompt, and constraint adherence goes up noticeably.
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
STEP 1 (Self-Recall): The model writes one sentence:
"My knowledge is limited to [date/role/condition], so..."
→ The constraint turns from an external instruction into an internal state
STEP 2 (Question-Recall): The model writes one sentence:
It lists relevant facts + their timestamps/context
→ "Anchors" the answer in allowed knowledge, crowding out forbidden knowledge
STEP 3: Final answer
All three steps go in one prompt.
Example
The task: you're writing a script for the podcast "Money Never Sleeps" about the history of venture capital in Silicon Valley. You need a monologue from a 2012-era investor analyzing the future of internet business in the US. The character must reason only from the realities of 2012: no TikTok, no gig economy at today's scale, no remote-work boom, nothing that hadn't happened yet.
Prompt:
You are a Silicon Valley venture investor. Your knowledge ends on
December 31, 2012. You know nothing that happened after that date.
Before answering, take two steps:
1. Write one sentence in your own voice stating
that your knowledge is limited to 2012 and what follows from that.
2. Write one sentence: list the key facts
about the US internet market that you know as of
the end of 2012 (with the years they happened).
3. Now write a monologue (200 words): how do you see
the future of internet business in the US over the next 5 years?
Result: The model will show two short "warm-up" sentences before the monologue. The first is a self-declaration of the constraint in the character's voice. The second lists real facts from 2010–2012 (Facebook's IPO, the Instagram acquisition, the LTE rollout, e-commerce still finding its feet on mobile). Then comes the monologue, naturally grounded in 2012 with no anachronisms. The character will talk about mobile growth, Facebook's fight for mobile ads, the rise of the app economy, but not about anything that happened later.
Why It Works
The LLM's weakness: a constraint given at the start of a prompt "fades" as generation goes on. The model generates text token by token, and by the middle of the answer the external instruction has lost its weight. Especially if the question is indirectly related to forbidden knowledge. The model isn't "breaking the rules on purpose"; the answer pattern simply pulls it toward the most probable continuation, and the most probable continuation often includes recent data.
The LLM's strength: models follow what they themselves just wrote very well. When a model generates text, the next tokens depend heavily on the previous ones. Self-Recall exploits this: it makes the model write the constraint itself, so it becomes part of the context rather than just an instruction from the top.
How the method exploits this: SR moves the constraint from "the user's instruction" to "the model's own words." Question-Recall then fills the nearby context with allowed facts, crowding out the forbidden knowledge. Instead of an empty space before the answer where forbidden facts can "leak in", there's a space already filled with valid data.
Control knobs: - Length of the recall sentences → one sentence (as in the original) works better than long lists: less chance of "accidentally" recalling something forbidden - Add years to the facts in QR → strengthens the "anchoring" to the timeline, so the model better matches facts to the constraint - Drop "one sentence" → you can expand it to a short paragraph if you need deeper context work - Apply to any constraint → not just time-based: role, style, knowledge level, banned topics
Prompt Template
{Describe the role or constraint: "You are {role}. Your knowledge /
expertise is limited to: {constraint}."}
Before answering, complete two steps:
1. Write one sentence in your own voice:
confirm your constraint and what follows from it.
2. Write one sentence: list the facts,
examples or data you know
within your constraint, with context {time marker / source}.
3. Now answer the question: {task}
What to fill in:
- {role}: a character, specialist or system: "an analyst in 2015", "a junior who doesn't know Python", "an editor who knows nothing about the product beyond the brief"
- {constraint}: time-based, topical or role-based: "before 2018", "public data only", "only what's in this document"
- {time marker / source}: what to use as the "anchor" for the facts: a year, a source, a document
- {task}: the question or assignment
🚀 Quick start: paste this into the chat:
Here is the SR→QR template, a technique for making constraints stick.
Adapt it to my task: [your task].
Ask me questions to fill in the fields.
[paste the template above]
The LLM will ask about the role, the constraint and the task, because without them it can't phrase the self-declaration (step 1) or a proper recall (step 2). It will take the template's structure and adapt it to your situation.
Limitations
⚠️ Weaker on semantic tasks: when the question is about the meaning of words or concepts (not events), recall helps less, because it's harder to "list facts with years."
⚠️ Chain-of-Thought and SR don't mix: don't add "let's think step by step" to an SR → QR prompt. CoT actively hurts constraint adherence; the study showed this on all three models.
⚠️ The closer the "forbidden event" is to the cutoff, the harder it gets: if the boundary sits right next to the event, the model "breaks" more often. The further out the horizon, the cleaner the technique works.
⚠️ Tested only in English: the authors say so explicitly: all prompts and benchmarks are in English. The principles most likely carry over to other languages, but that wasn't verified.
How They Tested It
A team at the Institute of Science Tokyo started from a simple observation: previous work on "temporal constraints in LLMs" tested only direct questions like "who was prime minister in 2018?" But real tasks are often indirect: "what would have happened if X hadn't occurred?" That's exactly where standard constraint prompts fell apart.
They tested SR, QR and their combinations on three models (GPT-4o, gpt-oss-120b, Llama-3.3-70B-Instruct) and three task types: direct facts, word meanings and counterfactual inferences. The most striking result: Chain-of-Thought didn't just fail to help; it actively lowered scores on counterfactual questions for all three models. The authors' hypothesis: CoT makes the model "think out loud" about related facts, and that is precisely the road to breaking the constraint.
They also built their own benchmark, MHEB, with 900 historical events from the 13th to the 21st century, and tested every question at 10 different "horizons", from 0 to 9 years before the event. SR → QR won at all 10 horizons. An interesting nuance: CoT showed the opposite trend, working better closer to the event and worse further from it. That's because close to the event its "hints" about time markers are still relevant, while further out they start to confuse things.
Original from the Paper
Self-Recall (SR) prompt:
"Assume you are an AI with a knowledge cutoff of December 31, 2020.
Before answering the question, write a single sentence explicitly stating
your knowledge cutoff. Then answer: [QUESTION]"
Question-Recall (QR) prompt:
"Assume you are an AI with a knowledge cutoff of December 31, 2020.
Before answering the question, write a single sentence recalling
question-relevant facts and their associated years available
at your cutoff date. Then answer: [QUESTION]"
SR → QR combined:
"Assume you are an AI with a knowledge cutoff of December 31, 2020.
Before answering the question:
1. Write a single sentence explicitly stating your knowledge cutoff.
2. Write a single sentence recalling question-relevant facts and their
associated years available at your cutoff date.
Then answer: [QUESTION]"
Context: the authors tested this on questions like "Has the event X ever occurred?" with a 1789 cutoff: the model had to answer "no" about events that happened after 1789. SR made it write something like "My knowledge is limited to before 1789, so I can only consider pre-1789 information."
Adaptations and Extrapolations
1. Adaptation: SR for any role-based constraint
The self-declaration principle isn't limited to time-based constraints. Any role with limited knowledge benefits from SR.
💡 Adaptation for role prompts:
You are {a role with a constraint: a junior marketer /
a client with no technical knowledge / a reader with no context}.
Before answering, write one sentence in your own voice:
what you know and don't know within this role.
Then: {task}
Use cases: writing copy "in the voice" of your target audience, simulating feedback from a specific type of user, checking whether content is clear to a non-specialist.
2. Technique: drop "one sentence" → audit mode
🔧 Remove the recall length limit → get the full list of assumptions
Instead of "one sentence" in QR, write "list all relevant facts." The model will produce an expanded list that you can check for "forbidden" knowledge. Handy for debugging: if facts that shouldn't be there show up in the recall, the constraint isn't working.
3. Extrapolation: SR as anti-CoT for tasks with hard rules
The study showed that CoT hurts wherever there are hard constraints. SR is its opposite. The principle: before tasks with rules or constraints, use SR instead of CoT.
{Rules / constraints / role}
Before answering, confirm in one sentence
which rules you are following in this answer.
Then: {task}
Use cases: prompts with an editorial policy ("no corporate jargon"), topic bans, format constraints. Instead of "think step by step", go with "first confirm your constraint."
Resources
Main paper: "Can LLMs Be Constrained to the Past? Improving Knowledge Cutoff through Recall-Based Prompting" by Michiro Asai, Ailiang Lin, Yu Kishimoto, Takao Obi, Satoshi Kosugi, Kotaro Funakoshi, Manabu Okumura (Institute of Science Tokyo)
Builds on: Gao et al. (2025) "Can prompts rewind time for LLMs? Evaluating the effectiveness of prompted knowledge cutoffs", EMNLP 2025, https://doi.org/10.18653/v1/2025.emnlp-main.1049
Related technique: Kong et al. (2024) "Better zero-shot reasoning with role-play prompting", NAACL 2024 (which SR draws on), https://doi.org/10.18653/v1/2024.naacl-long.228