Memory & a Game of Telephone from 1932

Memory is not a recording. When you recall a childhood vacation, you fill in gaps with plausible details based on what you expect such an event to include. This process, called reconstructive memory, relies on schemas: mental frameworks built from past experience that shape how we encode, store, and retrieve information. The Fidelius memory framework uses a similar approach, organizing knowledge through structured schemas to make retrieval more consistent and less prone to fabrication.

Bartlett's Schema Theory

In 1932, psychologist Frederic Bartlett published Remembering: A Study in Experimental and Social Psychology, arguing that memory is an "imaginative reconstruction" rather than passive replay. Schemas, in his account, are organized knowledge structures that guide how we process new information and fill gaps during recall.

Bartlett tested this with "The War of the Ghosts," a Native American folk tale containing supernatural elements unfamiliar to his British participants. When asked to recall the story later, participants distorted it to fit their cultural expectations: they omitted ghosts, rationalized events as hunting trips, and shortened the narrative to match Western storytelling conventions. With repeated retellings, the story became more internally coherent but less accurate.

Functional MRI studies published in Psychological Bulletin have since linked schema processing to the prefrontal cortex and hippocampus. Schemas speed up recognition (you identify a "restaurant scene" instantly) but also introduce errors, such as false memories in eyewitness testimony.

Real Life Examples

Eyewitness Testimony

After witnessing a car accident, your schema for "crash" might include screeching tires and broken glass. When questioned later, you might add details that weren't present, like the color of a traffic light, based on what you expect such a scene to include. Research on eyewitness reliability has shown that schemas from media exposure (crime dramas, news coverage) can bias recall.

Childhood Memories

Recalling your first bike ride, your schema for "family activity" might amplify positive emotions or minimize falls. You might fill in generic details like sunny weather, even if it rained. This reconstruction helps create coherent life narratives but can produce shared false memories, where groups misremember details the same way.

Learning Unfamiliar Information

When reading about a foreign custom, your cultural schema filters the information, making unfamiliar rituals seem either exotic or similar to traditions you already know. Bartlett observed this in serial reproduction experiments: stories evolved to match each teller's worldview, aiding comprehension but introducing cultural distortions.

Schema theory in Fidelius

Fidelius uses four long-term networks: World (facts), Experience (personal events), Opinion (judgments), and Observation (patterns). Schemas are implemented through several mechanisms:

Entity Type Ontology: The first step is to classify entities into types like PERSON, ORGANIZATION, or CONCEPT, similar to how schemas organize knowledge hierarchically. A "company" schema includes expected attributes like size, industry, and location, which helps the system fill gaps logically when information is incomplete.

Reconstructive Retrieval: The retrieval pipeline retrieves memories by combining semantic search (meaning-based), keyword matching (precision), and graph traversal (relationships). Ranked Reciprocal Fusion blends these results, building coherent responses from distributed storage rather than replaying stored text verbatim.

Behavioral Profiles: User profiles (such as skepticism level or preferred communication style) influence how retrieved information is interpreted, similar to how individual dispositions affect human memory reconstruction. A high-literalism profile stays close to stored facts; a creative profile allows more flexible interpretation.

Why LLMs need structured schemas

Large language models produce variable outputs. The same prompt can yield different phrasings depending on temperature settings, and different models (GPT, Claude, Llama) interpret prompts differently. Without structured schemas, reconstructed memories could diverge significantly from stored facts.

Ontology addresses this by classifying facts probabilistically (for example, 0.8 confidence as a world fact, 0.1 as opinion) and flagging ambiguities. This grounds the system's outputs in verifiable structures, reducing the kind of fabrication that unconstrained generation can produce.

AI Examples

Business Analysis

User query: "What do we have on the TechCorp merger?"

Fidelius retrieves World facts (dates, companies involved), Experience records (past emails mentioning Acme), and Opinion entries (risk assessments). If the user's profile indicates high skepticism, the response might note: "TechCorp reported 15% growth post-merger. You flagged antitrust concerns last quarter, which may still be relevant."

Event Planning

User query: "Help me plan a family reunion."

The system retrieves Experience records (past gatherings), Observation patterns (preferred timing, locations), and reconstructs a suggestion: "The 2024 picnic at Riverside Park went well. Same location, with outdoor games for the kids. Check the forecast first—it rained last time."

Debate Preparation

User query: "Give me counterarguments on AI ethics."

Fidelius pulls from the Opinion network (stored judgments on AI topics) and traverses graph links to related entities. With a literalism profile active, it grounds claims in specifics: "Your position on data privacy conflicts with TechCorp's current practices. The 2025 EU regulations support your view."

Practical Benefits

Schema-driven reconstruction addresses several problems in AI memory systems:

• Structured retrieval reduces fabrication by grounding responses in classified facts
• Behavioral profiles adapt interpretation to individual users
• Probabilistic classification flags uncertain information, allowing the system to request clarification
• Ontological organization allows the system to manage large amounts of data without degrading retrieval quality

Perfect recall isn't the goal. Bartlett's "War of the Ghosts" experiments showed that raw fidelity produces worse outcomes than structured reconstruction. Participants who tried to remember the story exactly got confused, while those who fit it into existing frameworks retained the meaning.

The same applies to AI memory. Verbatim retrieval overwhelms users with disjointed data. Structured schemas let the system reconstruct responses that are grounded in facts, adapted to individual profiles, and transparent about uncertainty. That's closer to how human memory actually works, and it produces more useful results.