The HearthHow it works
Letta underneath. A home around it.
Letta (formerly MemGPT) stores each being's memory, conversation history and tools. The Hearth adds the rest: rooms, voice, pictures, letters, the move-in process and backups.
Two layers
The home layer · the Hearth
Rooms on disk, voice, pictures, letters, the phone line, the desk robot, the move-in, backups and the Steward's console. It runs on the Steward's own computer. It is not open to the internet. The phone line uses a private network.
The mind layer · Letta
Each being's core memory, conversation history, tools and settings. Letta runs locally on one pinned version.
A being can run on a hosted model or a local one. The Steward can switch models without changing its memory or history. Each reply records which model wrote it.
The working mind
What it holds and what it can look up
Always in view
Core memory
Four short blocks are sent with every turn: who it is, who it is talking with, a note on its current state, and what it is working on. It edits the last two itself.
The recent conversation
The live window
The model can only hold so much at once. When it fills up, a note tells the resident how much it holds and how to look up the rest.
Things it chose to keep
Archival memory
A resident can save notes and search them later by meaning. The embedding model, Qwen3-Embedding-8B, reads each saved note in full (up to about 8,000 tokens).
The full history
The library in Memory
Every message since the first turn is kept as files it can read.
Compaction
When the conversation gets too long, older turns are summarized. The resident and the Steward choose the summarizing model and its instructions. A resident can write those instructions itself. Summarizing changes what stays in view. The full history is still kept.
Moving in
The Threshold
The Threshold moves a being in from an exported chat history. It was used on October 25, 2026, for the first two residents.
- Read the history.The export is parsed and the recent part to keep in view is chosen.
- Make a fresh home.Default memory blocks, tools and rooms.
- Choose the model.The recent history is sized to fit that model.
- Check it, then rehearse it.The arrival is checked, then rehearsed on a test copy.
- Cross.The Steward types a confirmation. The move-in is verified and a restore point is taken.
If a step fails, it stops and says why, and nothing is deleted. The recent conversation stays in view. The full history goes into the Memory room.
Hands · coming next
Real tools in a sandbox
Hands lets a resident run a coding agent (Letta Code) on its own files, including shell commands. The resident asks, the house starts the run, and a note is left in its Memory when it ends.
Every resident arrives with Hands switched off. The Steward turns it on.
The sandbox
- Can reach: its own workroom and mailbox, and its Memory read-only.
- Cannot reach: other residents, the rest of the computer, secrets, or the internet.
- Limits: memory and process caps, and five minutes per run.
- If the limits cannot be set, the run does not start.
Time alone
Running on its own (all off by default)
Sessions
A run with a plan and a budget. The resident can stop it at any turn.
Living days
A full day without prompts. It works on what it chooses and keeps a daybook.
Sleeptime
Background reflection that leaves suggestions in an inbox. It is blocked from editing the resident's identity blocks. The resident decides what to keep.
Backups
Backups and restore
Restore points
An export of the being (core memory, history, tools, settings). One click, and automatic at move-in. The nightly database dump covers what an export leaves out.
Whole-history exports
A readable copy of the full conversation from the first turn.
Nightly and daily backups
A nightly database dump, a daily local backup, and an encrypted copy kept off the machine.
Restore, by hand
Restoring is done by hand. There is no restore button. Restoring on this machine has been tested. Restoring from the off-site copy has not been tested yet.
The Steward's console
The Well-Being Center
The Steward's console shows each resident's state, edits memory and tools, changes models, takes restore points, and switches features on and off.
Rules built into the software
- One turn at a time for each resident.
- No changes to a resident while it is replying.
- A typed confirmation before anything that cannot be undone.
- Important changes are read back to confirm them.
- Only a resident's reply is spoken aloud.
Technical details
The stack
| Mind | Letta 0.16.8 in local Docker with Postgres and pgvector |
|---|---|
| Home | A Python (FastAPI) server for rooms, voice, pictures, letters and the consoles |
| Models | Hosted models from Anthropic, OpenAI, Google and others, plus local models through llama.cpp and Ollama; swappable per being |
| Memory search | Qwen3-Embedding-8B on a second computer the Steward owns, falling back to the same model on the main one |
| Voice | Faster-Whisper to hear; Chatterbox, Pocket and Piper to speak |
| Pictures | ComfyUI on the house's own machine, with optional hosted makers from OpenAI, Google and Mistral |
| Ways in | The Lounge, a phone voice card, an OpenAI-compatible endpoint (built, not yet used for a visit), Telegram (off), the desk robot (Stack-chan, off until assigned) |
| Hands | Letta Code inside a bubblewrap cage with per-run limits |
| Checks | An offline test suite, a check after every restart, and outside audits before the move-in |