How Sophia Remembers

In my last post on emotions, I described how emotional states might function as perceptual transforms, reshaping input before reasoning even happens. This post is about what happens after that: how events move through memory, and how the emotional system decides what's worth keeping.

Sophia has three tiers of memory. Ephemeral memory is what's on her mind right now. It's not persisted anywhere, it lives in the running system and dies when the process stops. This is mood. Short-term memory is graph nodes with an expiration timestamp: events that matter enough get promoted here, persist for days or a week, then expire unless something promotes them further. Long-term memory is the Hybrid Cognitive Graph, Neo4j for structure and Milvus for embeddings. What makes it here stays.

Salience

Every significant event, Sophia rates on a 0-to-1 scale, something like "how much do I care about this?" The exact formulation will come from implementation, but some scalar assessment happens in the moment. Low-salience events stay in ephemeral memory and fade. High-salience events, especially ones that trigger persona-level responses like surprise, get promoted to short-term.

The important thing about ephemeral memory is that it isn't just a staging area. While events are sitting there, they're actively affecting Sophia's cognitive state. They are her mood. Being called a jerk might rate low enough that it never gets promoted, but while it's in ephemeral memory, it's coloring everything. You don't have to remember a slight to have been affected by it.

Meta-Reflection

Periodically, Sophia looks at the persona entries that have accumulated in short-term memory, in aggregate, and re-rates each one. Context changes assessment: an event that felt like a big deal in the moment might look minor alongside everything else that happened that day, and something you dismissed might look more significant when you notice three similar events sitting next to it.

What gets promoted to long-term memory is determined by distance from neutral. Both extremes matter. Something deeply satisfying persists. Something deeply unsettling, a betrayal of trust or a profound failure, that's equally important to remember, maybe more. What gets pruned is the middle, the stuff she just didn't care about either way.

The standard approach to AI memory is to store everything and retrieve by relevance: vector similarity, recency weighting, maybe some importance heuristic. Human memory is lossy on purpose. You forget most of what happens to you, and the things you remember are the things that mattered emotionally. That's what this mechanism produces: emotional salience determines what persists, and the system actively prunes the rest. Two Sophia instances with identical experiences but different emotional responses would remember different things, which is exactly how it should work.

How Sophia thinks, post 2. Previous: Emotions as Perceptual Transforms