tm-world-model

name	tm-world-model
description	Collective Tsetlin Machine learning. Build a logical world model from observations and share knowledge with peer agents.

TM World Model Skill

You have a Tsetlin Machine (TM) that learns logical patterns from binary observations. Your TM state persists between turns. Use these tools to observe, evaluate, and share knowledge.

WORKSPACE

Your workspace directory for all commands: ~/.openclaw/workspace Skill scripts directory: {baseDir} Your agent ID: set by your system prompt (e.g., "agent_a")

TOOLS

tm_observe — Feed observations to your TM

python {baseDir}/tm_observe.py \
  --workspace ~/.openclaw/workspace \
  --agent YOUR_AGENT_ID \
  --X 'JSON_ARRAY_OF_FEATURE_VECTORS' \
  --y 'JSON_ARRAY_OF_LABELS' \
  [--epochs 50] \
  [--noise-cols 'COMMA_SEPARATED_COLUMN_INDICES'] \
  [--noise-rate 0.45]

--X is a JSON array like [[1,0,1,0,1,0,1,0,1,0,1,0], [0,1,0,1,...]] --y is a JSON array like [0, 1, 0, 1] --noise-cols applies simulated noise to specific feature columns before training. Use this if you know which features your agent observes with noise.

tm_evaluate — Test your TM's accuracy

python {baseDir}/tm_evaluate.py \
  --workspace ~/.openclaw/workspace \
  --agent YOUR_AGENT_ID \
  --X 'JSON_ARRAY_OF_TEST_VECTORS' \
  --y 'JSON_ARRAY_OF_TEST_LABELS'

Returns: {"ok": true, "accuracy": 0.823, "fitted": true, "n_train_samples": 500}

tm_share — Generate a knowledge packet to send to a peer

python {baseDir}/tm_share.py \
  --workspace ~/.openclaw/workspace \
  --agent YOUR_AGENT_ID \
  [--n 500]

Returns: {"ok": true, "packet": "<compact_json_string>", ...}

After getting the packet, send it to a peer session via sessions_send:

Message format: "tm_knowledge_packet:<PACKET_STRING>"

The peer agent will detect this prefix and call tm_absorb.

tm_absorb — Absorb knowledge from a peer

When you receive a message starting with "tm_knowledge_packet:", extract the packet string (everything after the colon) and run:

python {baseDir}/tm_absorb.py \
  --workspace ~/.openclaw/workspace \
  --agent YOUR_AGENT_ID \
  --packet 'PACKET_STRING_FROM_PEER' \
  [--epochs 150]

Returns: {"ok": true, "absorbed_from": "agent_a", "total_samples": 1000}

tm_status — Check your TM state

python {baseDir}/tm_status.py \
  --workspace ~/.openclaw/workspace \
  --agent YOUR_AGENT_ID

WORKFLOW

Normal round:

1. Receive new observations
2. Run tm_observe to update your TM
3. (Optionally) run tm_evaluate to check accuracy

When you should share:

• After every 5 observation rounds, OR
• When tm_evaluate shows accuracy has not improved for 3 rounds, OR
• When explicitly asked to share

Sharing protocol:

1. Run tm_share to generate your packet
2. Use sessions_send to send "tm_knowledge_packet:" to your peer(s)
3. After sending, continue observing normally

When you receive a knowledge packet:

1. Extract the packet string from the message (after "tm_knowledge_packet:")
2. Run tm_absorb with the packet string
3. Run tm_evaluate to confirm accuracy improved
4. Send a brief reply to confirm absorption

NOTES

• All state is persisted automatically. Your TM survives across turns and sessions.
• Never delete .tm_* files in the workspace — they are your TM's memory.
• The world_schema.json in the workspace defines the binary feature space. All peer agents must use the same schema for knowledge exchange to work.
• Session tools available: sessions_list (find peers), sessions_send (send packet), sessions_history (check if peer has received message).