There's a particular small jolt you get mid-conversation, when something the other person just said doesn't square with something they said a while back. Wait — didn't you tell me last month this hadn't started? Half the time you let it go, because you're not sure, because calling it out feels aggressive, because your memory of last month is fuzzy enough that maybe you're the one who's wrong. So the gap sits there, unspoken, and the conversation keeps running on two different sets of facts.
Around the fire, the circle remembered the earlier telling, so a story that changed got noticed without anyone having to be the bad guy — it was just collective memory doing its job. Contradictions is that collective memory, for your recordings.
The thing a per-meeting summary cannot do.
Almost every meeting tool is scoped to a single recording. It can summarize this call beautifully and has no idea that call ever happened. That scoping makes one whole category of insight structurally impossible: anything that requires comparing what was said now against what was said before. A tool that forgets every meeting the moment it ends can never catch a story that changed between meetings, because it only ever sees one.
Bonfiyah's whole architecture points the other way. Speakers persist across your library, so the same person in March and in May is understood to be the same person. Once that's true, reasoning across recordings stops being science fiction and becomes the obvious next move. Contradictions is what that move looks like in practice, and it sits directly on top of the cross-recording voice ID that ties a person's recordings together in the first place.
What Contradictions actually surfaces.
It watches for the places where a person's account drifts over time, and brings them to your attention with both quotes in hand. Status that moved is the obvious one: “basically done” today, “hasn't started” last month. Either the work moved fast or the story did, and it's worth knowing which. Numbers that shifted are close behind — the budget that was “around forty” in one conversation and “fifty, easily” in another, or the timeline that quietly grew two weeks every time it came up.
Then there are commitments that softened: a firm “yes, we'll have it by the 15th” that becomes “we're aiming for the 15th” becomes “the 15th was always a stretch.” Watching a commitment erode in slow motion is its own kind of early warning. And finally facts that contradict — two statements that simply can't both be true, said far enough apart that nobody would catch it from memory alone.
In every case, Contradictions shows you both moments, both exact quotes, both timestamps, side by side — a before/after evidentiary pair. It doesn't editorialize and it doesn't accuse. It puts the two tellings next to each other and lets you decide what they mean. Crucially, it only raises a pair when the two statements actually conflict: rather than matching keywords, it runs natural-language inference — does the later statement logically contradict the earlier one? — so “we're shipping by June 30” and “we never committed to June” register as a conflict even though they share no words, and two statements that merely sound similar don't trip a false alarm.
And because every commitment eventually either holds or doesn't, the same engine quietly keeps score. Over enough recordings, each recurring person earns a follow-through score — how often their word holds up — built from the kept-versus-missed record rather than a gut feeling, and surfaced on their People Memory profile. It's calibration, not condemnation: you learn whose “Friday” means Friday without keeping a private ledger in your head.
Tone matters, and the tone is on purpose.
A feature like this could very easily feel hostile — a lie detector bolted to your conversations, primed to catch people out. That's deliberately not what this is, and the framing reflects it.
Most discrepancies are completely innocent. People misremember. Situations genuinely change. “Hasn't started” in March was true in March; “basically done” in May is true in May; nobody lied, the work just happened in between. The job of Contradictions isn't to assume bad faith — it's to make sure both people are working from the same set of facts, so the conversation doesn't quietly fork into two incompatible versions of reality. Far more often than it catches anyone in anything, it catches you misremembering, and saves you from confidently asserting something wrong.
It's named Contradictions, not Lies, for exactly that reason. The point is a shared, accurate record — not a weapon.
Where it earns its keep.
Long negotiations are where it shows up first — terms that drift across weeks of back-and-forth. When the other side's “final number” is the third different final number, having all three quoted and dated changes the conversation. Vendor and contractor relationships are another natural fit: “it's on track” said cheerfully across five status calls, against the one call where the timeline first slipped. The pattern is the signal.
Beyond that, it earns its keep anywhere high-stakes and multi-session — care coordination across specialists, a board's evolving account of the same metric, any place where the same topic gets revisited and the story has room to wander.
Contradictions leans on the same foundation as People Memory: durable speaker identity and reasoning across your whole library. One keeps a living picture of each person; the other notices when that picture stops being internally consistent.
A note on respect and consent.
Reasoning across everything someone has said is powerful, so it lives behind the same guardrails as the rest of Bonfiyah. The conversations are yours, in your library, and the consent tooling rides alongside them in every tier — surfacing two-party rules, capturing verbal consent, keeping an exportable log. Contradictions works on recordings you were part of and chose to keep. It's a memory aid for your own conversations, not surveillance of anyone else's.
Try it across a few conversations.
Contradictions needs history to work — it's comparing now against before, so it has nothing to compare until there's a before. Record several conversations with the same people over a week or two and let the Pro AI layer run. The first time it quietly surfaces a “didn't you say…” you'd otherwise have second-guessed yourself out of raising, you'll understand what it's for.
Bonfiyah records on your device and transcribes in the cloud; the provider transcript is deleted immediately and your original audio auto-deletes from our servers within seven days. Bonfiyah is a universal Apple app — iPhone, iPad, and Mac, synced over iCloud — so the library Contradictions reasons over follows you across every device you record on. You can start free — 120 minutes a month, add Pro AI at intro pricing, and point it at a week of real conversations.
The circle always remembered the earlier telling. Now your recordings do too.
— Richard
