conscious-agent 2.2.0

Self-referential structure-learning system — agents that learn by inhabiting worlds

conscious-agent

A computational implementation of Hoffman's Conscious Realism.

Build self-referential agents that learn by inhabiting worlds — constructing internal models of both their environment and themselves.

For AI coding assistants: A SKILL.md file lives in .context/SKILL.md with patterns for complex use cases (multi-agent networks, crystal projection, live data feeding, debugging). opencode and compatible tools load it automatically.

from conscious_agent import ConsciousAgent
from conscious_agent.worlds import CoinTossWorld

world = CoinTossWorld(n_coins=4)
agent = ConsciousAgent(world=world, agent_id="my_agent")
outputs = agent.run(n_steps=1000)
print(f'"I" locked: {agent.is_i_locked}')

Installation

pip install numpy scipy       # core dependencies
pip install conscious-agent   # once published

Or from source:

cd hoffman-agents-python
pip install -e .

Quick Start

Single agent in a coin-toss world

from conscious_agent import ConsciousAgent
from conscious_agent.worlds import CoinTossWorld

world = CoinTossWorld(n_coins=3)
agent = ConsciousAgent(agent_id="coin_agent", world=world)

for _ in range(500):
    output = agent.step()
    if output.i_locked:
        print(f"I locked at step {output.step}")
        break

Custom Markov world

from conscious_agent import ConsciousAgent, WorldBuilder
import numpy as np

data = np.random.rand(500, 3)
world = (WorldBuilder()
    .add_feature("temp", normalization="minmax", n_bins=4)
    .add_feature("humidity", normalization="minmax", n_bins=4)
    .add_feature("pressure", normalization="minmax", n_bins=4)
    .build(data))

agent = ConsciousAgent(agent_id="weather_agent", world=world)
outputs = agent.run(n_steps=1000)

Combine two agents

from conscious_agent import combine

a = ConsciousAgent(agent_id="agent_a", world=world)
b = ConsciousAgent(agent_id="agent_b", world=world)
a.run(500)
b.run(500)

combined = combine(a, b)
print(f"Combined agent: {combined.agent_id}, level: {combined.cycle_level}")

Multi-agent network

from conscious_agent import AgentNetwork

network = AgentNetwork(n_agents=10, seed=42)
states = network.run(n_generations=100)
print(f"Avg prediction error: {network.avg_prediction_error():.3f}")

Save and load

from conscious_agent.io import save_agent, load_agent, clone_agent

path = save_agent(agent, "./souls")
loaded = load_agent(path)

cloned = clone_agent(agent, "experiment_clone")

Public API

# Core classes
from conscious_agent import ConsciousAgent, World, WorldBuilder
from conscious_agent import SimpleWorld, ExperienceSpace

# World factories
from conscious_agent.worlds import CoinTossWorld, build_world_from_dataframe

# IO
from conscious_agent.io import save_agent, load_agent, clone_agent, load_latest

# Multi-agent
from conscious_agent import AgentNetwork, combine

# Core components (for advanced use)
from conscious_agent import (
    TraceBuffer, TraceEvent, ExperienceTrie, MetaTrie,
    SelfTokenState, ExperienceLexicon, strange_loop_score,
)

# v2.0 — Agent mode control
agent.set_mode("frozen")        # 'learning', 'frozen', 'debug'
agent.thaw()                    # back to learning mode
agent.refreeze()                # back to frozen

# v2.0 — Memory & lifecycle
agent.clear_memory()            # reset trace buffer + counters, preserve trie
agent.inject_observation(world_state)  # push new data mid-run

# v2.0 — Metrics & introspection
agent.metrics                   # { prediction_error, i_locked, loop_depth, ... }
network.get_metrics()           # { agent_count, mean_prediction_error, i_lock_rate }
network.get_agent_metrics(id)   # individual agent's metrics
trie.get_stats()                # { node_count, max_depth, mean_visit_count, ... }
trie.export_nodes(3)            # all paths with visit_count >= 3
trie.get_dominant_paths(5)      # top 5 most-visited paths

# v2.0 — Batch stepping
network.step_all(world_state)   # step all agents with same world state
network.agent_list              # agents as an ordered list

# v2.0 — Action space
output.action_distribution      # { token: probability, ... }
agent.set_allowable_tokens({"I", "notice"})  # constrain output

# v2.0 — Composition
combine(a1, a2, a3)             # n-ary combination (3+ agents)

# v2.0 — TraceBuffer
trace_buffer.resize(100)         # dynamic window resizing

Self-Awareness

This library provides four self-awareness mechanisms, three built-in and one optional:

Mechanism	Type	What it does
MetaTrie	Built-in (implicit)	Models the agent's own trace buffer patterns — a hidden self-model
SelfTokenState ("I")	Built-in (implicit)	Tracks identity stability; locks on meta-trie convergence
strangeLoopScore	Built-in (explicit)	Measures self-referential depth in output tokens
SelfWorld	Optional wrapper	Injects agent's internal metrics into its perception stream

SelfWorld

SelfWorld is a world wrapper that lets the agent perceive its own internal state alongside external data. The agent's trie learns transitions over composite states of (world + self).

from conscious_agent.worlds import SelfWorld

inner = SimpleWorld(n_states=10)
agent = ConsciousAgent(
    agent_id="self_aware",
    world=SelfWorld(inner, lambda a: {
        "sp": a.experience.self_token.stationary_prob,
        "pe": a.mean_prediction_error,
    }),
)
agent.run(n_steps=1000)

Each step, the agent's WorldState contains both 'world' and 'self' sequences. The agent discovers patterns like "when my prediction error is high and the world shows pattern X, the next state tends to be Y."

→ Full philosophical architecture: docs/SELF_AWARENESS.md

How It Works

Every ConsciousAgent has an experience space — four interconnected structures:

1. TraceBuffer — short-term memory: the last N state transitions
2. ExperienceTrie — long-term world model: compressed prefix tree over observed state sequences
3. MetaTrie — self-model: a second trie over the agent's own trace buffer snapshots (thinking about thinking)
4. SelfTokenState ("I") — identity: the dominant meta-state that forms a stable attractor

The agent cycles through perception (observe world → update trie) → meta-observation (observe self → update meta-trie) → decision (generate output tokens via ergodic Markov chain).

When the meta-trie's stationary distribution converges on a single meta-state, the "I" locks — the agent has formed a stable identity.

Requirements

• Python 3.10+
• numpy >= 1.24
• scipy >= 1.10

License

MIT