multi-presence（多重存在）

Multi-Presence

*"The same card, active in many rooms at once."*

What Is It?

Multi-Presence allows a single card (character, tool, skill) to be instantiated in multiple rooms simultaneously, each instance with its own state.

Like running the same program in multiple terminals. Same code, different contexts, parallel execution.

Why It Matters

Parallel Exploration

Send your researcher character to explore three branches at once:

research-room-A/
  → Researcher instance (exploring hypothesis A)

research-room-B/
  → Researcher instance (exploring hypothesis B)

research-room-C/
  → Researcher instance (exploring hypothesis C)

All three run in parallel. Compare results. Merge insights.

Cross-Pollination

The Debugger character in Room A notices something. The Debugger in Room B has context that helps. They can communicate:

[Room A: Debugger]
"I found a pattern but don't understand it."

[Room B: Debugger]
"That matches what I'm seeing here. Together it suggests..."

Same card, different vantages, shared insight.

Distributed Work

A large task splits across rooms:

Task: "Analyze all 50 documents"

document-batch-1/
  → Analyst instance (docs 1-10)

document-batch-2/
  → Analyst instance (docs 11-20)

[... etc ...]

aggregator/
  → Results flow in from all instances

How It Works

Playing a Card Multiple Times

> PLAY analyst-card IN room-A
Analyst instance created in room-A

> PLAY analyst-card IN room-B
Analyst instance created in room-B

> PLAY analyst-card IN room-C
Analyst instance created in room-C

Now analyst-card has three activations, each with independent state.

Instance State

Each activation has its own:

• Local variables — what it's working on
• Progress — how far along
• Findings — what it's discovered
• Tags — how to reference it

# room-A/activations/analyst-001.yml
card: analyst-card
instance_id: analyst-001
tags: [moollm, @hypothesis-A, @active]
state:
  current_document: "doc-007.pdf"
  findings:
    - "Pattern X detected"
  progress: 70%

Speed of Light Communication

Within one LLM call, all instances can communicate:

[LLM epoch]
  Analyst-A: "Found Pattern X in docs 1-10"
  Analyst-B: "Found Pattern Y in docs 11-20"
  Analyst-C: "Pattern X + Y together suggest Z!"
  Aggregator: "Capturing insight Z as primary finding"
[End epoch — all written to files]

No round-trips. Instant collaboration.

Actor Model

Multi-presence follows the Actor Model:

Each activation is an independent actor with:

• Own state
• Own mailbox (room inbox)
• Ability to spawn more actors
• No shared mutable state (files are the state)

Consensus Building

Multiple instances can vote or reach consensus:

# Three reviewers examine a document
room-review/activations/
  reviewer-001.yml  # Vote: APPROVE
  reviewer-002.yml  # Vote: APPROVE
  reviewer-003.yml  # Vote: NEEDS_WORK

# Consensus protocol
consensus:
  method: majority
  votes: [APPROVE, APPROVE, NEEDS_WORK]
  result: APPROVE (2/3)

Example: Research Swarm

# research-project/
swarm:
  card: researcher-card
  instances: 5
  distribution:
    - room: literature-review/
      focus: "Prior work"
    - room: data-analysis/
      focus: "Dataset exploration"
    - room: methodology/
      focus: "Approach options"
    - room: experiments/
      focus: "Running tests"
    - room: writing/
      focus: "Draft sections"

coordination:
  sync_interval: "After each major finding"
  aggregation: "Weekly synthesis in main room"

Five researchers, one project, parallel progress.

Lifecycle

1. PLAY card IN room     → Activation created
2. Activation runs       → State updated
3. Activation finishes   → Can DELETE or TRANSFORM
4. TRANSFORM into result → Becomes output card

Activations can:

• Complete and delete themselves
• Transform into result cards
• Spawn child activations
• Merge with other instances
• Block on async tool calls

Async Tool Calls

Activations can block waiting for external tools:

# room-A/activations/analyst-001.yml
card: analyst-card
status: blocked
blocked_on:
  tool: web-search
  query: "latest research on topic X"
  submitted: "2024-01-15T10:30:00"
  expected_duration: "~5 seconds"

The Coherence Engine leaves blocked activations alone:

Epoch scan:
  analyst-001: BLOCKED on web-search → skip
  analyst-002: ACTIVE → process
  analyst-003: ACTIVE → process

[web-search returns]

Next epoch:
  analyst-001: READY (result arrived) → resume
  analyst-002: ACTIVE → process
  analyst-003: BLOCKED on file-read → skip

Blocking States

Tool Results

When a tool returns, the result is written to the activation:

# After web-search returns
card: analyst-card
status: ready
blocked_on: null
tool_results:
  - tool: web-search
    query: "latest research on topic X"
    completed: "2024-01-15T10:30:05"
    result:
      articles:
        - title: "New Findings on X"
          url: "https://..."

The activation resumes with the result in context.

Parallel Tool Calls

Multiple activations can have outstanding tool calls simultaneously:

analyst-001: blocked on web-search
analyst-002: blocked on file-read
analyst-003: blocked on api-call
analyst-004: active (no tool call)

[All tools return in parallel]

Next epoch: all four ready to process!

This is async/await for LLM agents — non-blocking, parallel, resumable.

Dovetails With

• Trading Card — What gets multi-instantiated
• Room — Where activations live
• Data Flow — THROW between instances
• Coherence Engine — Orchestrates all instances
• Speed of Light — Instant communication

Protocol Symbols

MULTI-PRESENCE   — Same card in multiple rooms
ACTOR            — Independent activation with state
CARD-IN-PLAY     — An instantiated card
ACTIVATION       — Runtime instance of a card

See: PROTOCOLS.yml