Schema Formation

Definition

Schema formation is the process by which commonalities recurring across many separate experiences accumulate into a single superordinate knowledge structure that no individual experience contains (⁠Gilboa & Marlatte, 2017⁠). What distinguishes it from memory for an event is that the resulting structure is general rather than specific: its elements stand for classes of thing rather than particular instances, and it is built across episodes rather than laid down in one (⁠Gilboa & Marlatte, 2017⁠). Two subprocesses are conventionally named. Assimilation fits a new experience into the structure already held; accommodation alters the structure to take account of an experience that does not fit (⁠Gilboa & Marlatte, 2017⁠). The process is slow for two reasons: the structure being extracted is a property of the whole set of experiences rather than of any one of them, so it can only be estimated by aggregating over many samples, and changing the structure rapidly to fit one new experience overwrites what is already held (⁠McClelland et al., 1995⁠).

Scope and contested boundaries: The neuroscience literature uses "schema" loosely, at times for any prior associative network that influences new processing. Gilboa and Marlatte (2017) adopt that inclusive usage deliberately, treating narratives and event gists, concepts and categories, and statistical regularities as forms of prior knowledge that resemble schemas and may be their components. Whether schemas proper differ from these qualitatively or only in degree is stated there to be unknown, and the same review calls for more precise definitions. This row covers the formation of superordinate structure across episodes; the use of an already-formed structure on new input belongs in Schema Application.

Neural Substrate

The human evidence on where schemas are represented comes from lesion work and functional imaging. Meta-analysis of imaging contrasts places schema reinstatement in the ventromedial prefrontal cortex, bilateral anterior temporal cortex, the temporoparietal junction including the angular gyrus, the hippocampus, and posterior midline cortex (⁠Gilboa & Marlatte, 2017⁠). Evidence bearing on formation rather than use is thinner and does not converge. Participants trained across nine months on object-location paired associates showed a shift from hippocampal to vmPFC representation over the first night, but by three months retrieval of those same overlearned associations engaged the ventrolateral prefrontal cortex with the anterior temporal lobe and angular gyrus, with no vmPFC involvement, a pattern closer to semantic networks than to schema networks; a shorter rule-learning study found vmPFC involvement at 24 hours (⁠Gilboa & Marlatte, 2017⁠). The review treats this as unresolved, and notes that how schemas are themselves constructed and accommodated remains the least studied aspect of the topic, since most work addresses the assimilation of new information into schemas already held.

The claim that schema-consistent material consolidates rapidly and ceases to depend on the hippocampus rests on rodent work and has not been demonstrated in humans. Rats trained across roughly a month of sessions on flavor-place paired associates in a fixed arena acquired two new associations in a single trial each and retained them after near-complete (approximately 90 percent) hippocampal lesions made 48 hours later, whereas lesions made 3 hours after learning left performance at chance (⁠Tse et al., 2007⁠). Attempts to show comparably rapid hippocampus-independent learning in humans, using fast-mapping paradigms in amnesic patients, have proved difficult to replicate (⁠Kumaran et al., 2016⁠). McClelland et al. (1995) is not a source of neural data here. It supplies the theoretical account of why extraction of structure across episodes has to be gradual, and the constraint it identified was subsequently amended in light of the rodent findings: the rate of neocortical learning is now characterised as dependent on prior knowledge rather than slow as such (⁠Kumaran et al., 2016⁠).

In humans, schema formation accumulates commonalities across many separate episodes into a superordinate structure that no single episode contains, over months, through a fast hippocampal system feeding a slow neocortical one (⁠Gilboa & Marlatte, 2017⁠; ⁠McClelland et al., 1995⁠). Artificial networks do the same thing in their weights and in the same order, acquiring broad distinctions before fine ones because dimensions with stronger statistical structure are learned faster (Saxe et al., 2019), and showing the same dependence on prior knowledge when facts consistent with what a model already holds are fitted quickly while facts that are not are fitted slowly (Gekhman et al., 2024), but no benchmark evaluates either as schema formation and no human baseline shares units with them.

• Weights: the numerical connection strengths between units in a neural network, adjusted during training, and the place where anything learned from training is held (McClelland et al., 1995).

How Far It Has Been Measured

Where It Lives in the System

01 Weights

What Has Been Measured

• Training: five pretrained language models (from BERT to MPT) almost always rated possible events as more likely than impossible ones across 1,215 sentence pairs but were less consistent for likely versus unlikely events, showing that generalized event knowledge partly emerges from training text; no human memory comparison was made (⁠Kauf et al., 2023⁠).
• Saved memory: Agent Workflow Memory has a GPT-4-based web agent distil reusable routines from its past tasks and store them for later use, raising task success on the WebArena benchmark from 23.5% to 35.5%; this is schema-like induction under another name, with no human comparison (⁠Wang et al., 2024⁠).
• Saved memory: generative agents periodically condense stored observations into higher-level 'reflections' that guide later behaviour, and removing reflection made behaviour less believable in ablation tests, but the abstraction itself was not scored against people (⁠Park et al., 2023⁠).
• Conversation: a 2026 preprint ran four current AI models on OpenLock, a causal-structure transfer task first used with 80 people, giving models a written description of a previously solved environment in their context window (what the model can see in the current conversation); no study was found giving AI a classic schema-formation paradigm such as prototype distortion or the Tse schema task (⁠Xiang et al., 2026⁠).
• People who had experienced the same causal structure before cut attempts from 20.66 (SD 9.09) to 13.85 (SD 10.00), d = 0.71, and cut attempts to their first solution from 8.97 to 4.99, whereas the best model, Gemini-3-Flash, improved only from 9.08 to 7.33 attempts and no model significantly cut first-solution attempts (⁠Xiang et al., 2026⁠).
• Transfer benefit varied widely across systems: 33% for people versus 19.3% for Gemini-3-Flash, 17.9% for Claude-4.5-Sonnet, 10.5% for DeepSeek-V3.2 and -10.5% (worse after prior experience) for GPT-5.2 (⁠Xiang et al., 2026⁠).
• The comparison is approximate: people played a visual version in 2018 and had to remember earlier environments, while models played a text version with the prior solutions still visible and had lower baseline attempt counts; because the uneven test favoured the models yet they still showed less use of prior structure, the below-human result counts, though it rests on one preprint with no direct statistical test between people and models (⁠Xiang et al., 2026⁠).