Memory Model

Persona organizes user memory into three typed classes: Episode, Psyche, and Goal. But what makes this model different is not just the types—it is how they connect through time.

​

Memory Types

All memories are stored in a graph database with embeddings for vector similarity search.

​

Time and Causality

Traditional memory systems categorize information into semantic, episodic, procedural, and other types borrowed from cognitive science. These categories are useful for describing human cognition, but they are fundamentally static. They treat memory as compartments to be searched independently. Human experience is not compartmentalized. It flows. One moment causes the next. A preference discovered today retroactively explains behavior from last month. A goal abandoned changes the meaning of past efforts. Memory is not a filing cabinet—it is a river with currents that flow in both directions. Causality in Persona is not just about provenance or walking backward to find sources. It is about understanding how memory evolves. When episodes are linked temporally, the system can observe patterns of change. A psyche trait like “anxious about public speaking” might appear in Episode 5, weaken in Episode 12 after a successful presentation, and disappear entirely by Episode 30. Without temporal structure, these are three contradictory facts. With temporal structure, they are a story of growth. This is what we mean by integration and consolidation. New information does not simply accumulate—it contextualizes what came before. When you ingest a conversation where the user says “I used to hate mornings, but now I prefer them,” the system can trace this evolution. The old psyche trait is not deleted; it is marked as superseded, with a temporal link showing when and why the change occurred. Causality also enables smarter retrieval. When you ask about the user’s preferences, the system does not just return the most recent statement. It understands the trajectory. It knows whether a preference is stable, evolving, or recently reversed. This temporal awareness transforms retrieval from “what did they say?” to “who are they becoming?”

​

Retrofitting

The most powerful consequence of temporal linking is retrofitting. When new data arrives—from a different source, at a different time—Persona does not simply append it to a list. It integrates the new information into the existing graph, creating connections to episodes that already exist. Imagine ingesting a year of calendar events after months of chat history. Traditional systems would treat these as separate corpora. Persona links calendar events to the chat episodes that surrounded them. The meeting you mentioned in March now connects to the calendar invite, the follow-up email, and the goal you set afterward. Context that was scattered across platforms becomes a unified narrative. This retrofitting happens automatically through temporal metadata and semantic linking. When an episode’s timestamp falls between two existing episodes, it is inserted into the chain. When its content relates to existing psyche traits, those connections are formed. The graph self-organizes around time and meaning.

​

Beyond Semantic and Episodic

The standard cognitive science taxonomy—semantic, episodic, procedural—was designed to explain how humans remember. It was not designed to build systems that remember for humans. Semantic memory stores facts without context. Episodic memory stores events without identity. Procedural memory stores skills without intent. These categories work for academic research, but they fail for personalization because they compartmentalize what should flow together. Persona’s model is different. Episode is not just “an event”—it is a link in a causal chain. Psyche is not just “a fact about the user”—it is a crystallized trait that always surfaces during retrieval. Goal is not just “a task”—it is a stateful intention that persists until completion. The pillars are different because the purpose is different. We are not describing human memory. We are building a living system that evolves alongside a human life.

​

Graph Relationships

Unlike traditional graph schemas that enforce rigid edge types, Persona treats relationships as fluid semantic descriptors. While the system maintains a strict temporal backbone using PREVIOUS and NEXT links, all other relationships are generated by the LLM based on context. This allows the graph to capture the nuance of human association that rigid types often miss. For example, a connection between a Psyche node and an Episode might be labeled derived_from, but a connection between two Episodes might be contradicts, reinforces, sparked_idea_for, or supersedes. This flexibility is central to the philosophy of the Memetic Organism. Just as thoughts in a human mind are connected by an infinite variety of associations, the edges in Persona’s graph adapt to describe exactly how two memories relate to each other. We provide the LLM with guidance and purpose, but we do not constrain it to a fixed enum of relationship types.