Schemas: How Prior Knowledge Quietly Distorts Every Memory

You see what you expect to see

You walk into a restaurant you have never visited before. You know there will be a host, then a menu, then drinks, then food, then a check. You know roughly where the bathrooms will be. You know how to read the body language of a waiter. None of this comes from this specific restaurant. It comes from the schema you have built up across hundreds of restaurant visits, and that schema is shaping every memory you form right now without you noticing.

The biology

Frederic Bartlett's 1932 work, including the famous "War of the Ghosts" study, established schema theory. Bartlett asked British students to read a Native American folktale, then repeatedly recall it over weeks and months. The recalls drifted systematically: unfamiliar elements got reshaped, replaced, or omitted, until the story was a more conventional Anglicized narrative. The students were not being lazy; they were running the story through their existing schemas, and the schemas won.

Schemas operate before, during, and after encoding:

• Before: they shape what you pay attention to in the first place.
• During: they organize new information into pre-existing structures.
• After: they distort retrieval, filling in plausible details that were never actually there.

Two related theories extend the idea:

• Scripts (Schank and Abelson, 1977) represent stereotyped event sequences: ordering food, going to the doctor, attending a funeral. Each script has expected slots and roles.
• Piaget's assimilation versus accommodation: assimilation fits new data into existing schemas. Accommodation modifies the schema to fit new data that does not match. Both processes are constantly at work.

Critically: schema-consistent information is encoded easily but can be distorted. Schema-inconsistent information sometimes is remembered better (because it stands out), sometimes worse (because the schema "rounds it off" toward expectations). Eyewitness memory failures are often schema artifacts.

The technology

Schemas are explicit in modern memory systems:

• Graphiti supports both prescribed ontology (define entity and edge types upfront via Pydantic models) and learned ontology (let structure emerge from data). This directly mirrors the assimilation versus accommodation distinction.
• MemOS provides custom memory schema definitions through configurable MemCubes.
• Cognee grounds knowledge graphs in formal ontologies (OWL, SNOMED CT, FIBO) with fuzzy matching for entity-to-class mapping. Each schema steers all subsequent processing, just as a biological schema steers all subsequent encoding.
• Fine-tuned models embed domain schemas implicitly in their weights.
• System prompts function as schemas biasing all subsequent processing. "You are a customer support agent" creates a script that shapes what counts as relevant information for the rest of the conversation.
• AutoSchemaKG (2025) learns ontologies automatically from data, analogous to children constructing new schemas through experience.

The pattern is recognizable: a schema (ontology, prompt, fine-tuned weights) acts as a strong prior. Information that fits the schema is processed efficiently. Information that does not fit either gets reshaped to fit (assimilation) or, if the system is sophisticated enough, gets the schema updated (accommodation).

Where the gap is

Custom ontologies are available in production. The frontier is dynamic schema evolution: schemas that adapt and restructure based on incoming data without human governance. Most current systems lock the ontology at design time. AutoSchemaKG and similar work are early but not yet production-default.

The deeper concern is one Bartlett identified almost a century ago: schemas distort. A heavily schematized memory system will lose information that does not fit the schema. If your ontology says "Customer has Email" and the data also has "Customer has Discord username," the schema-bound system either drops it, awkwardly stuffs it into "Email," or silently fails. Schema accommodation, the human capacity to update the structure when the data demands it, is largely a human-in-the-loop process today.

Practical implication: be deliberate about your schema. A loose schema captures more raw information but loses retrieval precision. A tight schema gives crisp retrieval but distorts data that does not fit. Both Graphiti's hybrid approach and Cognee's ontology grounding are reasonable starting points. And expect to update the schema regularly as the data teaches you what you got wrong on day one.

Series footer

← Previous: Context-Dependent Retrieval · Series anchor · Next: Prospective Memory →