What Is an AI Memorial?

The category is young enough that implementations differ wildly, and the differences matter more than the label. Two products described identically can behave very differently depending on whether responses are anchored to real material or generated freely.

How a grounded AI memorial is built

1. Collection. Stories, interviews, letters, photographs with context, and audio. The quality ceiling of the memorial is set entirely here.
2. Structuring. Material is broken into passages and indexed so it can be retrieved by meaning rather than by filename.
3. Retrieval. When someone asks a question, the system finds the passages that genuinely relate to it.
4. Presentation. A language model phrases the retrieved material as a reply, in a manner consistent with how the person spoke.
5. Voice, optionally. With consented recordings, replies can be spoken in a familiar cadence rather than read as text.

The important step is the third one. A system that retrieves before it answers is constrained by the archive. A system that answers first is improvising, and improvisation about a dead person is exactly what families do not want.

What it can answer well

• Stories the person told, in something close to how they told them.
• Facts they documented: places, dates, names, and relationships.
• Values and advice they explicitly expressed, on subjects they addressed directly.
• Descriptions of photographs and objects, when context was captured with them.

What it cannot answer, and should decline

• Opinions on events after their death. There is no basis for one, and any answer is invention.
• Subjects they deliberately never discussed. Silence in an archive is information, not a gap to fill.
• Anything requiring judgement about a living person's current situation. It cannot know it.
• Precise wording of conversations that were never recorded.

How it differs from an online memorial page

A traditional memorial page is authored once and read passively. Its strengths are permanence, simplicity, and neutrality — everyone sees the same thing. An AI memorial is responsive: each visitor's questions lead somewhere different, and a grandchild can pursue what they specifically care about instead of reading someone else's summary. The trade-off is complexity and the need for care about accuracy. The full comparison is here.

Emotional considerations

Reactions differ genuinely and unpredictably. Some people find hearing a familiar voice steadying and return to it occasionally, the way they might reread a letter. Others find frequent interaction keeps a loss immediate rather than easing it. Research in this area is early and we are not going to pretend it is conclusive.

Practical guidance rather than clinical guidance: introduce it slowly, let each family member decide their own level of contact, and treat a decision to step away as entirely normal. If grief is interfering with daily functioning, that is a conversation for a qualified professional, not a product.

Consent and control

The cleanest situation is one where the person built the archive themselves and stated what they wanted done with it. Where a memorial is created after death by family, the standard should be higher: use documented material, avoid synthesising a voice without clear prior consent, and agree in advance who may access it. These questions are covered in the ethics of AI after death.

Where Eterna fits

Eterna supports both directions: a person building their own legacy while living, and a family preserving someone who has died. Personas draw on the memories, stories, and media in that person's archive rather than inventing a personality, memorials are private unless deliberately shared, and voice features require explicit consent and authorisation.

You can see the interaction model in the walk-through without creating anything. Our Acceptable Use Policy sets out what is not permitted, including impersonation and unauthorised voice use.