How the runtime works.
Most AI systems are stateless — they forget you the moment the session ends. Chalie is a runtime built on persistent memory. It holds context across sessions, reasons about intent through a deterministic router, and executes only with your approval.
What happens when you say something
Every message you send travels through these steps. Most complete in about one to two seconds.
You send a message
Plain text. No commands, no syntax. Chalie reads exactly what you wrote.
Memory loads in the background
Chalie pulls relevant memories — recent conversation, known facts, past episodes — in under 10 milliseconds. You don't feel the wait.
Intent is classified, not guessed
A fast mathematical router (not another AI) decides how to respond: answer it, ask a question, take an action, or simply acknowledge.
A response is prepared
The AI writes a response shaped by your conversation history, your known preferences, and the mode the router selected.
You stay in control
When Chalie wants to do something consequential — send a message, schedule something, run a tool — it asks first. Safe internal actions (searching memory, storing a fact) happen silently.
Memory is written for next time
After every exchange, a background process condenses the conversation into memories. Next time you speak, the context is already there.
How the runtime remembers
Memory is layered — just like how humans hold a current thought, a recent conversation, and a long-term impression all at once.
Right now
— current conversationChalie holds the last few turns of your conversation in working memory. When you reference something from a moment ago, it knows exactly what you mean.
Recent context
— summaries and factsAs conversation flows, Chalie compresses it into short summaries (gists) and atomic facts. These survive for hours — bridging across sessions on the same topic.
Long-term episodes
— narrative memoriesImportant exchanges become episodic memories — narrative units that include what happened, the emotional tone, and the outcome. These decay slowly over time, just like human memory.
Semantic knowledge
— learned conceptsOver time, Chalie extracts the concepts and relationships that matter to you — your values, recurring interests, preferred ways of working — building a semantic map of who you are.
Personal traits
— your identityFacts about you — your name, timezone, preferences, communication style — are stored as personal traits. They inform every interaction and fade naturally if they become outdated.
Lists and tasks
— deterministic stateShopping lists, to-dos, and chores never decay. They are stored with perfect fidelity and full history — Chalie always knows exactly what is on the list.
All memory layers are stored locally on your machine. Nothing is sent to an external server.
How the runtime decides
Most AI systems ask another AI "what should I do?" before deciding. That doubles the wait time and introduces unpredictability. Chalie does it differently.
How to respond — decided by maths
A scoring function reads signals from the conversation — how many facts are known, whether there's a question, whether this is a new topic — and picks the best response mode in about 5 milliseconds. No AI involved. Fully auditable.
What to say — decided by AI
Once the mode is chosen, the AI gets a focused, specific prompt: "write a clarifying question" or "write a direct answer." The mode narrows the scope, so the AI doesn't meander.
The four ways Chalie can respond
Give a direct, substantive answer. Used when context is warm and the question is clear.
Ask a question. Used when something is ambiguous or context is missing — Chalie won't guess.
Gather more information first — search memory, check a tool, run a calculation — then respond.
A short social reply. Greetings, thanks, and confirmations don't need a two-second AI call.
Your data stays yours
Privacy is designed in — not bolted on. These are the guarantees that can't be turned off.
Runs on your machine
Chalie runs locally. Your conversations, memories, and personal data never leave your device. There is no cloud database holding your history.
Zero telemetry
Chalie makes no external calls except to the AI provider you configure. No usage data, no analytics, no behaviour reporting. It cannot phone home.
Credentials are encrypted
Any API keys or external service credentials are encrypted before being stored. They are never exposed in plain text — not in logs, not in config files.
You hold the keys
You can inspect everything Chalie knows about you — autobiography, traits, memories, routing decisions — and delete what you choose. Transparency is not optional — it is built into the API.
Consequential actions require your approval
Safe actions like searching memory happen automatically. But anything consequential — sending a message, scheduling, running a tool — is verified by a critic and paused for your approval if needed.
Open source — inspectable by anyone
Every line of code is public. You don't have to trust our claims — you can verify them yourself, or have someone do it on your behalf.
Extending the runtime
The runtime extends through containerized embodiments — sandboxed modules covering search, research, Gmail, Google Calendar, Telegram, and more. New embodiments can be added without restarting.
Embodiments run in containers
Each embodiment runs inside a sandboxed container. It cannot read your files, access your memory, or touch other tools. If it misbehaves, it can't spread.
Results are checked before use
After a tool runs, a critic verifies the output before it influences Chalie's response. Consequential errors are flagged and escalated to you — not silently accepted.
Anyone can build one
Embodiments follow a simple module pattern. If you can write a script that reads stdin and writes stdout, you can build one. No SDK required. No proprietary format.
Ready to run the runtime?
The runtime runs on your machine. No account required. No data leaves your device until you configure an LLM provider.