A Memory Has Provenance: Preserve What Triggered It

A photograph on the table

An old photograph lies on a table. Someone recognises a building, then a neighbour, and recalls an event absent from the written archive. The valuable result is not only the story. We also need to know which image was shown, when it was shown, who asked the question and what the witness actually said.

Recording the answer alone removes part of its meaning. A memory, like a document, has provenance.

An archive can become a trigger

KAKENHI project 25K03415 at the University of Tsukuba runs from 2025 to 2030. It proposes matching residents with historical resources, collecting oral histories through robots and AI-assisted reminiscence, and later linking narratives to the resources that prompted them. These are research aims, not demonstrated outcomes.

Its predecessor, project 18H03491, built an immersive audiovisual prototype for up to four people to share local memories. The final report demonstrates the feasibility of a participatory workflow suitable for public libraries. It does not establish that robots outperform people or that AI produces more accurate memories.

The trigger belongs to provenance

A photograph is not merely an attachment. It participates in the situation in which the account is produced. The same is true of a map, deed, letter, object, place, audio excerpt or question. Tim Fawns’s research on photo-elicitation examines remembering as distributed across people, photographs and practices rather than treating the image as neutral.

Autobiographical memory is reconstructive, and retrieval cues influence access to it. This does not make testimony false. It means testimony should remain interpretable alongside its production context: what was presented, what was asked, what was answered and how certain the witness felt.

With AI, preserve the mediation too

If AI selects the next photograph, proposes a follow-up, rephrases a question, suggests a name or links an account to a document, that intervention must remain identifiable. AI may facilitate testimony; it must never become an invisible source of it.

Keep separate the historical source, a human question, an AI-proposed question, the witness’s words, transcription, later interpretation, automated inference and human validation. An algorithmic suggestion is never the witness’s memory.

What this changes for digital memory

A structured memory could retain the chain ‘trace → question → witness → narrative segment → place → consent → validation’. The point is not to keep more data, but to preserve our future ability to understand how information entered the memory.

That chain can distinguish documented, testified, inferred, contested and incomplete claims. It preserves contradictions and human control. For Mnemory Places, Legacy and Knowledge, provenance is a condition of responsible transmission.

Put it into practice

Step 1 — Free account. Begin without factual suggestions.

Step 2 — Present an identified trace: photograph, document, map, object or place.

Step 3 — Ask an open question: ‘What do you recognise?’, ‘What does this image bring back?’ or ‘Who might know more?’

Step 4 — Preserve the chain: trace, exact question, human or AI author, witness, date, recording segment, transcript, expressed confidence, consent and later validation.

Step 5 — Never erase corrections. Extend the history instead. Oral History Association guidance also stresses informed consent, interview preparation, preservation plans and, where possible, review before public release.

Preserve the conditions of emergence

Preserving testimony is not only recording words. A transmissible memory also keeps the conditions that allowed those words to emerge. Linking trace, question, mediation and response gives future readers the means to understand, challenge and transmit the account.

Philippe Contal