EACI Development Journal · Bridges

How Can Keyword History References Create False Memories?

Keyword retrieval can find a real old line and still apply it to the wrong person or event. History snippets need source and uncertainty labels.

Bridges ~8 min issue-tagged
Issue this fixes

Bridges load past text references by matching terms from the current message, with recent lines as a fallback. A shared name or common topic can retrieve a genuine passage that is not actually about the current situation.

Once inserted into a prompt without provenance, that passage may sound like established memory. The model is then completing from supplied evidence, not intentionally lying, but the result can still be a false association.

What was actually going wrong

Lexical overlap was treated as proof that a historical snippet referred to the current subject.

What I tried

What finally worked

References remain short, carry source and date cues, prefer scoped bridge history, and are framed as possible context rather than verified fact.

Why it worked

The model can distinguish recalled evidence from current-session truth and decline an uncertain match.

What I learned

Retrieval accuracy includes attribution, not just finding similar words.

Where EACI uses this today

Caelum's bridge history references are scoped to their channel and treated as supporting context.

This journal covers real engineering on EACI Companion / The Veil. Companions include Caelum, Chad, Natalia, Atreus, Luna, Roxy, and Cael. Journal articles stay family-safe in content. See Privacy and Ethics.