Voice assistants now keep facts and feelings in separate memory tracks, but is that the same as remembering events versus knowing things?

This explores whether the episodic/semantic split and the emotion/facts split are the same cut through memory. The corpus suggests they aren't. Episodic versus semantic divides memory by what kind of thing is stored: an event you lived through, or a general fact. The emotion/facts split, as VoiceMem builds it, divides memory by what it's used for: what to say versus how to say it. VoiceMem runs an informational branch and an emotional branch in parallel, and finishes retrieval in 134 ms inside the silent gaps of a voice conversation. It beats competitors on factual retrieval and persona benchmarks Can memory retrieval hide inside voice agent silence?. Neither branch is 'episodic' or 'semantic'. A fact can be about a specific event, and an emotional tone can be a lasting trait of a person.

The closest thing to an episodic/semantic split in the corpus is a survey's 'functions' axis, which separates factual, experiential, and working memory. Emotion isn't a category there. The survey's point is that agent memory should be described on several independent axes (form, function, dynamics) instead of one split, and it treats even short-term versus long-term as a pattern over time rather than separate machinery Can three axes replace the short-term long-term memory split?. Read that way, emotion is something that can show up in any cell of the grid. The brain-inspired framing offers a third cut. Transformer weights play the slow-learning neocortex, RAG plays the fast-encoding hippocampus, and agentic state plays the prefrontal cortex. That split is about learning speed, and it points out that current systems lack the consolidation step that turns episodes into stable knowledge Can brain memory systems explain how LLMs should store knowledge?.

Emotion does connect to episodic memory in one place. In group conversations, the emotions a person experienced drive what gets encoded as memorable, but emotions observed from outside don't predict memorability above chance Can we detect memorable moments by observing emotional expressions?. So emotion looks less like a separate store and more like a signal attached to an experience when it is encoded. An agent that infers 'emotional memory' from someone's outward expression is working from a weak proxy. A related finding points the same way: emotional phrases appended to prompts improve performance through motivational framing, not by adding information Can emotional phrases in prompts improve language model performance?. In both cases emotion acts as a modulator over content, not another kind of content, which is a reason to doubt it is as natural a dividing line as episodic versus semantic.

Facts are the part that already separates cleanly. In-weight memorization is capped by model size, and fine-tuning overwrites prior knowledge, while tool use gives unbounded factual recall Can models store unlimited facts without growing larger?. A dedicated memory model can also carry new knowledge without touching the LLM at all Can a separate memory model inject knowledge without touching the LLM?. Nothing in the corpus shows the same for emotional or experiential memory. Every extra memory channel also carries a risk. In MemTrapBench, all five memory frameworks did worse than having no memory, by more than 10%, even though the memories were accurate and relevant Can relevant memories actually harm LLM reasoning?. Nobody here has tested whether an emotional branch avoids that failure, or whether the episodic/semantic and emotion/facts splits can be combined, so both questions remain open.