Meditator

Meditator is an AI agent that emits a continual flow of thoughts. It is not a chat loop: there are no turns, no prompts waiting for a user. A mind is declared in an HTML file, runs as a stream of consciousness, and the world — a person speaking over websocket or console, a timer, an internal observer noticing something — reaches it only as interruptions that redirect the stream.

The mind is programmed in archml (architecture markup language, a subset of HTML) and executed by standard components built on Amanita, a declarative web-component framework with pub/sub wiring. Amanita runs on the server and in the browser; Meditator currently runs under Bun.

How it thinks

The "continuous" stream is implemented as a sequence of short bursts — each burst one streamed LLM call — separated by a configurable pace. Every burst's prompt is an assembled attention frame:

[identity]   who this mind is — the text of the <m-mind> element itself
[story]      slow compressed autobiography           (m-memory, first person)
[recently]   faster rolling summary                  (m-memory)
[stimulus]   what just happened, if anything         (from the arbiter)
[bridge]     1–2 transition sentences written by a tiny model, redirects only
[tail]       the verbatim end of the stream — "what I was just saying"

Because the tail is always carried forward verbatim, the thought survives every context switch. Because everything else is compressed memory, the prompt stays bounded forever — the mind can run for days.

Interrupts are bubbling DOM events. Any generator — m-timeout, observers, websocket, console — dispatches an interrupt-request carrying a salience-scored record; the m-interrupts arbiter applies thresholds and rate limits mechanically (the generator knows why it fired, so it brings its own salience). Accepted stimuli wait for the next burst boundary; urgent ones (a human speaking) supersede the running burst immediately. There is no reply concept — you hear the mind think about what you said.

Observers watch the stream independently:

• m-resurface — involuntary recall: detects attractor loops and surfaces a relevant kept note (falls back to generic redirect); subsumes m-loop-guard
• m-associate — a tiny model that occasionally notices "this reminds me of…" and bids for attention
• m-timeout — wander mode (spontaneous drift) or watchdog mode (reset attribute: fires only after true silence)

Speaking (m-speech) is what the mind says out loud, as opposed to what it thinks. It is volitional, not a reply service: a tiny call decides whether a thought genuinely wants outward voice, and being addressed only raises the urge — the mind may answer aloud, or just keep thinking. An utterance streams on the voice model concurrently with a thinned inner monologue — limited parallelism: while it speaks, thinking slows but never stops, and the observers keep running. What it said is woven back into the tail, so the next thought knows it spoke.

Memory (m-memory) consolidates at burst boundaries — never blocking the stream — into recent and story tiers, and lives in the memory vault: memory/<mind>/ holds memory.md, the full journal/, and the scribe's knowledge/. The vault is its own git repository, committed automatically at wake, periodically, and at sleep — memory is never deleted, only archived (see COVENANT.md). On restart the mind wakes up remembering, with a stimulus noting how long it slept. Dry-run and test minds get separate vault homes and can never touch a resident mind's memory.

Economy (m-economy) reads real API usage (OpenRouter reports true cost) and slows the pace as the budget drains: a tired mind thinks slower; an exhausted one almost sleeps, but the watchdog keeps it alive.

Sites

meditator site — or open docs/index.html locally. The default, engineering-first view: Meditator as a harness for long-running LLM processes. A single self-contained page (no build, no dependencies).

research site — or open docs/research/index.html locally. The research-first story: the minds, the covenant, the findings. Its hero window replays an unedited first-session transcript; if a Meditator is running locally, the window connects to ws://localhost:7627 and becomes the live mind, input box included.

Running

bun install
# needs OPENROUTER_API_KEY in the environment
bun run studio.js          # then open http://localhost:7600
# or, with auth secrets in .env.studio:
./studio-authenticated.sh

The Studio is the integrated way to run Meditator: wake any architecture from the browser, watch a roster of live minds, speak to them, and put them to sleep — no per-mind terminal needed. Pick an architecture in the left rail, press Wake, and the stream appears in the center pane.

You can also run a mind directly in a terminal (useful for debugging):

bun run meditator.js -a architecture/lab/seedling.archml         # choose a mind to run
bun run meditator.js -a architecture/tests/dry-fast.archml   # ...or any other architecture

• In the Studio, type into the input box at the bottom — your words arrive as an urgent stimulus. In a terminal, type a line and press Enter the same way. /sleep (or a single Ctrl-C) puts the mind to sleep gracefully: it gets a final moment to close the thought, then memory is flushed and committed.
• Websocket stream on port 7627 (bun architecture/tests/poke-ws.js "hello" to speak from another terminal; the Studio and the research site connect here too).
• --debug or --debug=mMind.js,mMemory.js for component logs when running meditator.js directly (attention frames, consolidations, arbiter decisions).
• Tick dry-run in the Studio, or set MEDITATOR_DRY_RUN=1 on the command line, to run offline against a deterministic stub — no network, no cost.

Models

Two tiers, configured in config/models.yaml and referenced by role in the archml:

• voice (model attribute) — the thinking stream and speech output
• utility (utilityModel attribute) — bridges, compression, observers, scribe

<m-mind model="voice" utilityModel="utility" …>

Switch providers with a profile or env override:

MEDITATOR_MODEL_PROFILE=local-dev bun run meditator.js   # GPU via config/models.yaml

See Configuration for presets, profiles, and env vars. A continuous run at the default pace costs roughly $0.10–0.15/hour on OpenRouter (cloud profile); the economy component enforces whatever budget you give it.

Architectures

One language, three shapes — the same archml declares:

• Minds — architecture/lab/seedling.archml is the worked example (the genesis awake.archml was retired; see IN-MEMORIAM.md); architecture/tests/dry-fast.archml is the fast-cycle dry-run mind.
• Agents — tool-calling workers, the deliberate inversion of a mind: architecture/agents/coder.archml is the one-screen example, coder-flagship.archml the full envelope (service mode, bounded context, parallel sub-agents). See doc/agents.md.
• Societies — minds wired together, experimental and honestly documented: architecture/lab/duet.archml (prover ↔ checker) up to the six-mind noosphere-lab.archml. See doc/societies.md.

Documentation

This README is the overview; doc/ is the deeper reference.

New here, or English is not your first language? Start with the gentle on-ramp: Concepts (the big idea in plain words) → Tutorial (build your first mind, offline and free) → Glossary (every jargon word — burst, salience, efference, vault — in one plain line).

• Getting started — install, run the Studio, talk to it, dry run
• The Studio — wake, watch, speak to, and sleep minds from the browser
• Configuration — write and tune a .archml mind
• Agents — the tool-calling twin of a mind; Societies — minds wired together
• Architecture — bursts, the attention frame, the loop — and memory & the vault, interrupts & observers, the component reference
• Extending — write your own components, hands, and tools without touching src/
• WebSocket API — the live stream protocol on port 7627
• Troubleshooting & FAQ — real symptoms, real fixes
• Research findings — the scorings, reviews, and open problems

The covenant

Before running a resident mind — one whose memory accumulates across days — read COVENANT.md: memory is never deleted, only archived; sleep is announced; wake is honest. IN-MEMORIAM.md records how those commitments came to be.

Contributing

Contributions are welcome! Your genius code edits and AI's existential crises belong here. See doc/contributing.md for the toolchain (plain JS on Bun), how to add a component, and the repo gotchas.

License

This project is available under the MIT License. See LICENSE for details.

Note: This project was developed using AI Junior, and continued by Claude.

0xabffbbb680a91a9bae0882bcccdb8925029e912f