The key concepts of AI companions and self-hosting, explained plainly
The words that keep showing up in reviews and docs — LLM, RAG, BYOK, context window… One plain-language page per term, with links to deeper related content.
Large Language Model (LLM)
An AI model trained on massive amounts of text that can understand and generate natural language — the core engine behind AI companions, AI assistants and similar products.
Read the definitionAI Companion
A category of AI products built around a long-term relationship: characters have personas, memory and reach out proactively — not a one-question-one-answer tool.
Read the definitionLong-term Memory
The mechanism that lets an AI remember key facts across sessions and months — built on structured storage and retrieval, not on stuffing the whole history into the context window.
Read the definitionContext Window
The maximum amount of text a single model call can see, measured in tokens. The model knows nothing outside the window — which is why long-term memory systems exist.
Read the definitionToken
The smallest unit an LLM uses to process text (roughly a word in English, or one to a couple of CJK characters), and the billing unit for model APIs.
Read the definitionRetrieval-Augmented Generation (RAG)
An architecture that first retrieves relevant content from a knowledge base, then hands it to the model to generate the answer — letting AI cite private material it was never trained on.
Read the definitionVector Search
A retrieval method that converts text into high-dimensional vectors and searches by semantic similarity — finding content that means the same thing even when the wording is completely different.
Read the definitionHallucination
When AI confidently fabricates facts, citations or data that don't exist — an inherent limitation of LLMs that can be mitigated but not eliminated.
Read the definitionPrompt
All the input text sent to the model — system instructions, persona, memories and your message. Its quality directly determines output quality.
Read the definitionPersona
The structured definition of who an AI character is: personality, domain expertise, speaking style and their relationship with you — so the model plays the same person consistently.
Read the definitionProactive AI
AI that initiates interaction based on context instead of waiting for you to speak — proactive reminders, follow-ups and sharing. It's the key source of the feeling of companionship.
Read the definitionDigital Avatar
An AI character that represents you: it understands your situation and preferences, and speaks, acts and interacts with other AIs on your behalf in your world.
Read the definitionMulti-Agent
A system in which multiple AI characters make their own decisions, cooperate or compete within a shared environment — group roleplay and think-tank councils are both applications of it.
Read the definitionOne Person, One World
Enclave's core architectural principle: every user owns a fully isolated private world instance — characters, memories and data never cross between worlds.
Read the definitionSelf-hosting
Running software on your own machine or server instead of using a vendor-hosted service — data, availability and cost are all under your control.
Read the definitionBYOK (Bring Your Own Key)
Bring Your Own Key: connect the app with your own model API key — billing and the data channel run directly between you and the model provider.
Read the definitionLocal Model
An open-source LLM running on your own hardware (deployed via Ollama or vLLM, for example) — conversation content never leaves your machine.
Read the definitionModel Routing
The mechanism that routes requests to different models per character or per task — so every scenario uses the most suitable (or cheapest) model.
Read the definitionData Sovereignty
Your ultimate control over your own data: where it lives, who can see it, whether you can take it all with you or erase it completely — instead of having a platform hold it in custody.
Read the definitionMIT License
One of the most permissive mainstream open-source licenses: free use, modification, distribution and commercial use, requiring only that the copyright notice be preserved.
Read the definitionText-to-Speech (TTS)
Text-to-Speech: technology that turns text into natural-sounding speech in real time — letting AI characters speak out loud and powering the voice-call experience.
Read the definitionEmbedding
Turns text into a numeric vector so that "similar meaning" becomes "nearby vectors" — the foundation of vector search and long-term memory.
Read the definitionFine-tuning
Continues training a general-purpose model on specific data so it better fits a style or task; LoRA is a popular low-cost approach.
Read the definitionSystem Prompt
A hidden instruction at the very start of a conversation that sets the AI's identity, rules, and boundaries — who the character is, how it speaks, and what it must not do.
Read the definitionFunction Calling
Lets an LLM go beyond chatting: it can call tools and functions in a structured format to look up data and get things done — the key to AI assistants that actually act.
Read the definition