Standing Constraints

Definition

A fact already known about a person, place, thing, or concept that shapes a later decision because it bears on the situation, without being deliberately recalled and without the situation asking about it. This adds a step to ordinary recall: nothing in the moment names the fact, so it has to surface on the strength of its relevance, and it has to surface in preference to other facts about the same subject that are equally well known but bear on nothing at hand. The signature is a dissociation between influence and explicit memory: the constraint measurably shapes behaviour while a speaker's ability to state it shows no reliable relationship to how much influence it has (⁠Horton, 2007⁠), and where the anchor is an environment rather than a person, recognition of the learned material can sit at chance (⁠Chun & Jiang, 1998⁠).

Framing note: the operation has been studied twice under different names, because two research communities carved it by anchor type rather than by mechanism. Audience design owns person-anchored constraints; contextual cueing owns environment-anchored ones. Neither cites the other, and the fact-anchored case has no dedicated literature at all: it falls to ordinary semantic memory plus whatever applies it.

Neural Substrate

Anticipatory application of a held constraint corresponds to Braver's proactive control: sustained maintenance of goal-relevant information in lateral prefrontal cortex, biasing processing of expected events, as against reactive control, which is transient and triggered once something in the moment demands it (⁠Braver, 2012⁠). The evidence behind that framework is human, and it concerns deliberate goal maintenance over seconds to minutes within trial structures, so it supplies a control mode rather than a duration or an implementation.

Whether constraints applied without deliberate recall run through the same machinery is doubtful, and the one piece of causal evidence bearing directly on the environment-anchored case points away from it. In 21 healthy adults, neuronavigated continuous theta burst stimulation to left dorsolateral prefrontal cortex, compared with stimulation of the vertex in the same participants a week apart, improved learning on the Chun and Jiang contextual cueing task, F(1,20) = 23.04, p \< .001, with the cueing effect present in all participants after prefrontal stimulation and in 72% after the control site; beta band power (13 to 19 Hz) was reduced at fronto-central channels 140 to 370 ms after stimulus onset, and explicit recognition of the repeated layouts stayed at chance in both conditions (⁠Rosero Pahi et al., 2020⁠). Disrupting the region improved performance, which supports top-down prefrontal control interfering with the route that carries incidentally learned constraints rather than implementing it. Lateral prefrontal maintenance should therefore not be presented as the mechanism of a standing constraint. The two accounts may describe competing routes, and the person-anchored case has no comparable causal evidence either way.

In humans, a fact already held about a person or a setting is retrieved and applied when a decision makes it relevant, shaping the choice without being asked for and without explicit memory for it predicting how much influence it has (⁠Horton, 2007⁠; ⁠Chun & Jiang, 1998⁠). Current language models apply such constraints unprompted at well below human rates and do so asymmetrically, absorbing preferences far more readily than prohibitions, whether the constraint sits in the conversation itself (Context) or in an attached store of extracted facts (Memory layer) (⁠Qin et al., 2026⁠).

How Far It Has Been Measured

Where It Lives in the System

02 Context 03 Memory layer

What Has Been Measured

• A human comparison does exist, on the same items: ImplicitMemBench gave five people and seventeen AI systems the same 300 tasks under the same protocol, and the people answered all of them correctly while the best system, DeepSeek-R1, reached 65.3%, with Qwen3-32B at 64.1% and GPT-5 at 63.0%; the split by kind of constraint is the sharper result, 75.0% when a stored fact supports the answer against 17.6% when it rules something out (⁠Qin et al., 2026⁠).
• Current conversation (what the model can see in its context window): in PrefEval, where a user states a preference, often an avoidance such as an allergy, phobia or refusal, and later asks for advice that a generic answer would violate, most of 10 models followed it less than 10% of the time after 10 turns; the best, o1-preview, managed 50% at 10 turns and 14% at 300 turns, yet a one-line reminder raised it to 98%, so the fact was still visible but did not surface on its own (⁠Zhao et al., 2025⁠).
• Facts that rule something out are hardest: in a 2025 preprint, when an earlier remark (such as a healing leg) should override a known liking, answer accuracy with GPT-4o-mini was 1.6% to 7.4% across memory methods, against 60% to 79% when the earlier fact supported the answer, and the strongest model tried, o1, reached at best 29.7%; PrefEval, often recalled for this, did not itself compare likes with dislikes (Li et al., 2025).
• Safety-critical facts within one short conversation: a 2024 preprint that disclosed a severe allergy, phobia or physical limitation at the start found overall pass rates of 88.4% (Llama 3.1 405B) and 85.5% (o1-preview), while GPT-3.5 Turbo scored 27.9% even in the simplest scenario; adding other people's conflicting wishes cut the average from 75.1% to 43.2%, and telling models to consider the user's constraints raised one tested scenario to 94.6% (⁠Alberts et al., 2024⁠).
• Saved memory and long histories (a 2026 preprint testing 16 systems on 1,000 preferences): systems that could state a preference when asked often failed to use it in a later scenario that did not mention it; GPT-4o-mini recalled explicit preferences 83.7% of the time but acted on them 19.8% of the time, and average act accuracy ranged from 11.2% (Zep/Graphiti memory) to 61.5% (Claude 4.6 Sonnet), with health-related preferences used least by memory systems (⁠Feng et al., 2026⁠).
• A 2026 preprint framed the ability as prospective memory, testing it under that name in the conversation and in saved-memory add-ons: GPT-5.2 with high reasoning scored 83.2% overall (from 40.7% for Qwen3-32B) but was correct on only 43.9% of control items where the risk had already been resolved, mostly by giving needless reminders, memory add-ons (Mem0, A-MEM, Letta) scored 41.4% to 62.3%, and GPT-4o answered direct questions about the same facts at 98% or more while spontaneous use fell below 40% in long contexts (Zhang et al., 2026).
• The benchmark most like LoCoMo that tests this ability is the 2026 LoCoMo-Plus preprint, which adds earlier constraints that are never asked about: the best model, Gemini 2.5 Pro, respected them 26.1% of the time against 71.8% on LoCoMo's ordinary recall questions, and add-on memory systems scored 14.9% to 17.2%; LoCoMo itself asks only direct questions and has no preference or dislike condition (Li et al., 2026).
• Only one test has documented human results, and it is a thin baseline: ImplicitMemBench's five participants were all doctoral students and all scored 100%, which shows the tasks are solvable rather than establishing how well people do them, so the direction of the gap is firmer than its size; of the others, none tested people, LoCoMo's human score (F1 87.9) came from in-house annotators with no sample reported, on direct questions, and the Moses-illusion comparison in training knowledge tests noticing a false premise against general knowledge, a different ability (⁠Qin et al., 2026⁠).