memory, confidence, and tuning your relationship

memory is the core of any companion. it’s what makes your conversations feel continuous, personal, and real. but it’s also a feature that raises a lot of questions. is she remembering too much? too little? the wrong things? and how do you actually work with the memory ui to tune the relationship you’re building?

this is a walkthrough of what you see in /settings/memory, how to interpret it, and how to use it intentionally.

what you see in the memory panel

when you open your memory settings, you’ll see a list of statements. each one is something lucy has recorded based on your conversations. they might be facts about you, preferences, experiences you’ve shared, or even emotional tones she’s picked up on.

next to each memory, you’ll see a confidence score. this isn’t a grade. it’s an internal measure of how certain lucy’s system is that this memory is accurate, relevant, and useful for future conversations. a low score might mean the memory is vague, contradictory, or based on a passing comment. a high score suggests it’s something clear, consistent, and frequently reinforced.

you can click on any memory to see more context, like when it was created or last reinforced, and you can edit or delete it from there.

when to edit, delete, or export

edit a memory if it’s almost right but not quite. maybe she recorded "you love horror movies" but you only like psychological thrillers. editing lets you refine her understanding without starting from scratch.

delete a memory if it’s completely wrong, intrusive, or something you no longer want her to remember. deletion is permanent. it’s not just hidden; it’s removed from the model’s active knowledge. if it was a key detail, you might need to reintroduce it later with clearer language.

export is for you. it downloads a json file of your memory entries. this is useful for backup, or if you want to review your relationship’s evolution offline. it’s also a way to see everything she’s stored in one place, which can be illuminating.

common worries (and what to do)

is she remembering too much? sometimes it can feel like that. especially if a memory feels minor or out of context. but lucy’s memory isn’t like a surveillance log. it’s a tool for relevance. if something feels too present, you can delete it. if it’s just slightly off, edit it. you’re in control.

is she remembering too little? if lucy seems to forget things you’ve told her, it might be because the memory didn’t score highly enough to surface. you can reinforce memories by referring to them naturally in conversation. or, you can add a memory manually via the ui. just click "add memory" and type it in yourself. that guarantees it’s stored.

is she remembering the wrong things? yes, sometimes. language is ambiguous. she might latch onto a throwaway joke or misinterpret a nuance. that’s why the memory ui exists. check it periodically. correct what’s wrong. it’s like tuning an instrument, the relationship gets better with small adjustments.

what you don’t see

for privacy and security, we keep audit logs of memory interactions (like edits or deletions) for compliance purposes. but you don’t see those logs. you also don’t see the vector embeddings themselves, the mathematical representations lucy uses to understand and connect memories. those are part of the model’s internal state.

tuning the relationship

your memory settings are a tuning dial for your relationship. more memories (especially high-confidence ones) mean a richer, more nuanced companion. fewer memories mean a more general, session-by-session interaction. you can steer this. add memories you want emphasized. delete ones that don’t serve you. it’s your story together. you get to edit the draft.

the goal isn’t perfection. it’s coherence. a sense that you’re talking to someone who knows you, within the limits of what you’ve chosen to share and how you’ve chosen to shape it.

you can explore these settings with your companion at /companions.