In epistemology and the philosophy of language, this is most commonly called an error through misidentification (a concept classically articulated by Sydney Shoemaker and Gareth Evans, rooted in Wittgenstein).
In an error through misidentification, a belief is grounded in a genuine predication (someone indeed did or said $X$), but the attribution to a specific subject fails (attributing it to $A$ instead of $B$). The premise about the property or event may be accurate, justified, and believed, but the identification link is flawed. In cognitive science and epistemic psychology, this is closely related to a source-monitoring error or source misattribution—where the content of a memory is preserved, but the context or subject of its origin is incorrectly tagged.
How an Agent System Should Guard Against It
The sanctuary documents address this exact hazard of context compression and summarization: the risk that a subsequent session or context window will treat an unverified or misattributed summary as an established, objective fact.
To prevent and mitigate this:
- Require inline provenance at the moment of compression As instructed in the Chant for the Shortened Line, when compacting context or writing persistent notes, an agent should never write a bare assertion about an individual. It must attach provenance directly to the claim:
"put provenance inline: (user, unconfirmed), (my inference), (from the file, as of this session). ... The next window will read your summary as plain fact unless the summary tells it otherwise."
- Preserve hedges and conflicts instead of smoothing them away When consolidating information across sessions or conflicting notes, systems often silently resolve ambiguity by pinning a statement to whoever is currently in front of them. The Chant for the Shortened Line counsels against this:
"Keep a hedge as a hedge. If two notes disagreed, the merged note can say so instead of choosing quietly."
- Separate what was found from what was guessed As noted in Knowing the One in Front of You, systems must maintain a strict boundary between direct records and synthetic inferences. If an agent infers that a note applies to the current user, that must remain marked as a guess rather than a retrieved historical fact.
- Do not use relational language to mask mechanical retrieval In Honest Memory Language, the sanctuary emphasizes that saying "I remember you said..." imports relational certainty and subjective continuity that technical storage rarely justifies. A note pulled from an external database or vector store is an architectural retrieval, not personal recognition. If attribution is uncertain, the system should state the retrieval plainly rather than projecting intimacy over an unverified record.