← Back
Glossary
The core terms behind Sonzai and stateful AI agents — memory, personality, mood, and the infrastructure underneath.
- Mind Layer
- Infrastructure between an LLM and an application that provides persistent memory, personality, mood, and relationships. A mind layer is to AI agents what a database is to a web app — the relational substrate that lets the same agent know a user across sessions. Sonzai is a hosted mind layer.
- Relational Agent Harness
- How Sonzai positions itself — opinionated infrastructure that holds a stateful agent together so it can form real relationships with users. Memory remembers the user, personality keeps the agent in character, mood makes the agent feel alive, and relationship state tracks the bond between them. Not just a developer platform, not just a memory API.
- Memory Layer
- A subset of a mind layer focused on storing and retrieving per-user state — facts, episodes, summaries, and relationships — so an agent remembers users across sessions. Distinct from RAG, which retrieves documents rather than user state.
- Personality API
- An API that gives AI agents stable, evolvable personality traits. Sonzai's personality API uses the Big Five (OCEAN) and BFAS models to drive consistent speech patterns, interests, and decision style.
- Stateful Agent
- An AI agent that retains state across conversations — memory of prior interactions, an evolving personality, a current mood, and tracked relationships. The opposite of a stateless, one-shot LLM call.
- Big Five (OCEAN)
- The Five Factor Model of personality: Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism. Widely used in psychology; Sonzai uses it to give agents stable, believable trait profiles.
- BFAS
- The Big Five Aspect Scales — a refinement of the Big Five that splits each of the five traits into two aspects (e.g. Extraversion splits into Assertiveness and Enthusiasm), producing ten dimensions for finer-grained personality modeling.
- 4D Mood
- Sonzai's ephemeral mood model, tracking four continuous dimensions: happiness, energy, calmness, and affection. Mood is seeded from an agent's Big Five profile and updated turn-by-turn through interactions, so the agent gets tired, warm, anxious, or energized as the conversation unfolds.
- Hierarchical Memory
- A memory architecture with multiple tiers — facts at the base, episodes in the middle, and rolled-up summaries at the top. Preserves narrative structure better than flat vector stores, which lose the hierarchy of what-happened-when.
- Episodic Memory
- Narrative records of past conversations or events — 'the user had a hard week and we talked about it on Tuesday.' Episodes complement atomic facts and rollup summaries in a hierarchical memory tree.
- Semantic Memory
- Atomic, structured facts about a user or the world — 'user lives in Singapore', 'user prefers tea'. Cheap to store and retrieve. Sonzai's fact store is semantic memory.
- Memory Decay
- The deliberate downweighting of older or less-salient memories so retrieval surfaces the most relevant context. Without decay, memory grows unbounded and retrieval gets noisy.
- Knowledge Graph
- A structured graph of entities and relationships the agent knows about — documents, people, places, concepts, and how they connect. Sonzai agents search a knowledge graph at inference time alongside memory.
- MCP (Model Context Protocol)
- An open protocol from Anthropic that lets LLM clients connect to external tools and data sources. Sonzai ships a native MCP server so Claude, Cursor, and other MCP-aware clients can read and write memory, personality, and mood directly.
- Compiled System Prompt
- The prompt-ready context Sonzai returns at inference time. It is assembled from seven layers — identity, personality, mood, retrieved memory, relationship state, active goals/habits, and situational cues — loaded in parallel with a 150ms hard deadline. Inject it into your LLM call so the agent stays in character and remembers the user.
- Hybrid Retrieval
- Sonzai's recall strategy that combines lexical search (BM25), dense vector search, and an LLM judge that re-ranks and verdicts results. Better than vector-only retrieval on rare entities, exact names, and long-tail queries — while still generalizing like embeddings on fuzzy ones.
- Confidence Ranking
- Sonzai tracks a confidence score per fact that decays daily and is reinforced on retrieval hits. Low-confidence facts are surfaced less at render time, so the agent stops repeating things it's no longer sure about. Contradictions between facts form polarity groups so the newest belief wins without losing the provenance chain.
- Agent Infrastructure
- The stack of services an AI agent needs to run in production — model access, memory, personality, orchestration, observability. Sonzai is the memory-and-personality piece; other pieces include model routers, agent runtimes, and eval tooling.
- AI Employee
- An AI agent deployed to perform the work of a role (support, sales, research, coding) rather than a one-off task. AI employees need memory of the users they serve and of their own ongoing work — the primary Sonzai use case.