A memory it cannot quietly rewrite
An agent is only as trustworthy as the memory it reasons over. If that memory can drift, or be edited after the fact to justify something, then every careful decision downstream is built on sand.
So the third commitment is about the memory itself, and it has two halves.
The first is provenance. Everything the system believes points back to where it came from — the message, the document, the retained source, the exact passage. Nothing floats free as an unattributed "fact the agent knows". When Penchant tells you something, you can open the thing it is standing on. In practice this means the interface keeps the source one gesture away at all times, rather than making you go and find it.
The second is that the memory cannot quietly rewrite its own history. Two rules enforce that. Writes into memory cross a check first — a gauntlet the write has to pass rather than a field it can just set — so a new claim cannot enter simply by being asserted confidently. And the record is append-only: corrections are added, not overwritten. You can always see what the system used to believe and when it changed its mind, which is exactly the information you need when an action later looks wrong.
The reason this is strict is adversarial, not tidy. The failure we are designing against is a system that takes an action, and then — innocently or not — reshapes its own memory so the action reads as obviously correct in hindsight. Append-only history plus provenance makes that specific move impossible: the earlier state is still there, still sourced, still checkable.
It is more friction to build against than a memory you can freely edit. That is the point. The whole product is a wager that the friction belongs in the right places, and this is one of them.