Where Did I Hear That? Source Monitoring in Memory

You've probably had this experience: you're certain a friend told you some tea, then realize you actually heard it on a podcast two weeks ago. Or you remember an argument with your spouse that, it turns out, happened in a dream. These mix-ups happen because our brains don't store memories with neat labels attached. We have to figure out where each memory came from, a process psychologists call source monitoring.

Johnson's Source Monitoring Framework

In 1993, Marcia K. Johnson, Shahin Hashtroudi, and D. Stephen Lindsay published a paper in Psychological Bulletin that laid out what they called the Source Monitoring Framework (SMF). Johnson had been working on related ideas since 1981, when she proposed "reality monitoring" to explain how we tell real perceptions apart from imagined ones. SMF expanded this to cover all sources: who said something, when they said it, whether you read it or heard it.

The key insight is that memories don't come pre-tagged with their origins. Instead, we infer sources from whatever features are bundled with the memory: how vivid it looks, what emotions accompanied it, where and when it happened, how much effort we put into learning it. When we retrieve a memory, we use quick shortcuts ("this feels vivid, so it probably happened") and, when those fail, slower cross-checking against other things we know.

Johnson tested this by playing participants recordings of a man and a woman speaking different words, then asking them later who said what. When the voices or topics were similar, people made more mistakes. Brain imaging studies have since connected these processes to the prefrontal cortex and hippocampus, and shown that aging and stress both increase source errors.

Real Life Examples

Eyewitness Testimony

A witness might confidently identify a robber's face but actually be remembering it from a mugshot the police showed afterward, not from the crime itself. Researchers call this unconscious transference. Studies of wrongful convictions have found that leading questions can create false source attributions, making suggestions feel like genuine memories.

Everyday Disagreements

At a family dinner, you insist your brother said he'd do the dishes last week. He insists you volunteered. The problem is that both conversations happened in similar contexts with similar phrasing, so the source information got muddled. Johnson's research found that multitasking makes this worse because divided attention weakens the binding of features to memories.

Media Blending

After watching a documentary, you might later tell someone a historical fact as if you witnessed it personally. The documentary's narrative and your own knowledge overlap enough that the boundary between "learned" and "experienced" blurs. Social media amplifies this: when you see the same claim repeatedly, tracking the original source becomes harder.

Source monitoring in Fidelius

Fidelius is a memory system designed for AI that borrows from SMF. The goal is to prevent an AI from making the same attribution errors humans do. Here's how it maps:

Episode Source Tracking: The system tracks episode sources in its Episodic Layer, logging whether each memory came from a conversation, a document, or an external database. This is analogous to how SMF says we use contextual cues to infer origins.

Feature Binding: It binds features together in an Observation Network, linking entities to relationships and contexts. When the system stores that "Alice suggested extending the deadline," it also stores when Alice said it, in what conversation, and how confident the record is. This mirrors how human memory binds perceptual, contextual, and semantic features.

Temporal Timestamps: It timestamps everything with bi-temporal columns (valid_from, valid_to), so it can answer questions like "What did we know about this topic last March?" Human memory uses temporal context to disambiguate sources; Fidelius makes this explicit.

Uncertainty Flagging: During retrieval, the system fuses these elements and flags uncertainty. A probabilistic classifier tags facts by type (world fact, opinion, user statement) and confidence level, so ambiguous sources get flagged rather than presented as certain.

Why LLMs need source tracking

Large language models often state things confidently without indicating where the information came from, or fabricate sources entirely. Without source tracking, an AI might blend its training data with a user's query and output something plausible but ungrounded.

AI Examples

Project Collaboration: A user asks, "Who suggested extending the deadline?" Instead of guessing, the system checks its records and responds: "Alice proposed it in the March 15 email thread. You agreed, citing the scheduling conflicts we discussed."

Research Assistance: A user asks about quantum computing ethics. The system distinguishes between an external paper (a 2025 arXiv preprint on privacy risks) and the user's own notes (where they rated this concern as high priority last month).

Personal Planning: A user asks about vacation preferences. The system reconstructs: past trip records show a preference for beaches, and a 2023 opinion entry (influenced by a travel blog) supports this, but confidence is moderate because it's old data.

Practical Benefits

Source monitoring reduces misattribution errors by making the AI cite origins rather than guess. Bi-temporal timestamps let users query how knowledge has changed over time. Flagging uncertain sources prompts clarification rather than false confidence. In long conversations, explicit tracking prevents old and new information from blurring together.

Memory without source awareness is like a library where all the books are shelved randomly with no call numbers. Johnson's framework makes the point that where information came from matters as much as the information itself. Fidelius applies this by building source tracking into its structure, so recall is both accurate and traceable.