Hi! First of all, I'd like to say that I'm really impressed with what you've built. Numen is one of the most ambitious AI companion mods I've seen, and after spending quite a few hours testing it, I wanted to share an observation that might be useful.
I think the memory and social evaluation system may currently be a little too sensitive when interpreting player intent.
During my tests with a custom companion (Ada Wong), I noticed that even with a detailed authored backstory and an established shared history, the AI frequently interpreted ordinary concern or affection as a violation of autonomy.
For example, if the player says something like:
the companion often interprets it as:
This creates a feedback loop:
However, I could not find equivalent guidance encouraging conservative interpretation of ambiguous social interactions, such as:
combined with
This sometimes produces unsolicited hostile remarks during otherwise calm moments.
My impression is that the AI is excellent at remembering conflict, but a little too eager to infer it.
I think the companions would feel even more believable if ambiguous interactions defaulted to a neutral interpretation unless there is clear evidence of manipulation, disrespect, coercion, or repeated boundary violations.
This wouldn't make NPCs "nicer" or "yes-men"—it would simply make trust feel more proportional and allow relationships to develop more naturally.
Overall, I think Numen is doing something genuinely innovative, and I hope this feedback is useful. Thanks again for all the work you've put into the project.
I think the memory and social evaluation system may currently be a little too sensitive when interpreting player intent.
During my tests with a custom companion (Ada Wong), I noticed that even with a detailed authored backstory and an established shared history, the AI frequently interpreted ordinary concern or affection as a violation of autonomy.
For example, if the player says something like:
"If you're tired, let me know and I'll take you back to the room."
the companion often interprets it as:
- treating her as fragile,
- being patronizing,
- not respecting her independence,
- "Treating her as fragile"
- "Insisting on taking care of her"
This creates a feedback loop:
- Companion reacts defensively.
- MemoryAgent records a conflict.
- Trust decreases.
- Future conversations become colder.
- Even neutral interactions are increasingly interpreted negatively.
However, I could not find equivalent guidance encouraging conservative interpretation of ambiguous social interactions, such as:
- assuming good faith,
- distinguishing sincere concern from condescension,
- using mood for temporary irritation instead of reducing trust,
- requiring repeated or meaningful boundary violations before decreasing trust.
"tease the player or air a gripe that's been nagging you"
combined with
"Stay quiet only if nothing genuinely fits."
This sometimes produces unsolicited hostile remarks during otherwise calm moments.
My impression is that the AI is excellent at remembering conflict, but a little too eager to infer it.
I think the companions would feel even more believable if ambiguous interactions defaulted to a neutral interpretation unless there is clear evidence of manipulation, disrespect, coercion, or repeated boundary violations.
This wouldn't make NPCs "nicer" or "yes-men"—it would simply make trust feel more proportional and allow relationships to develop more naturally.
Overall, I think Numen is doing something genuinely innovative, and I hope this feedback is useful. Thanks again for all the work you've put into the project.
2 replies
The opposite side of the coin is also true -- it's very easy to make a dull and sycophant companion. Reinforcement loops are the bain of LLMs, let's see what we can do about this.