Project Overview

Two layers live here:

• Graph tools: build/query a neuron graph from text (nn_builder.py, query.py, graph_api.py).
• Self-growing mind: an agent that mutates plugins/*.py to satisfy tasks, with stages, drives, tasks, web consults, reflections, and a diary.

Setup

Python 3 required. Install deps:

pip install -r requirements.txt

If using Postgres for graph ingestion, also install psycopg/psycopg2. Run a local LLaMA-compatible endpoint for reflections (config below).

Graph layer (optional)

• Build/lint: python3 -m py_compile nn_builder.py query.py graph_api.py
• Ingest to SQLite: python3 nn_builder.py --text-files path/* --sqlite-out graph.db --max-neurons 300000 --max-tokens-per-file 50000 --bridge-limit 50000
• Serve API: python3 graph_api.py --db graph.db --host 127.0.0.1 --port 8000
• Query: python3 query.py --db graph.db socket or curl "http://127.0.0.1:8000/ask?text=...".

Self-growing mind

Run the loop:

python3 -m manager.life_loop

Behaviors:

• Perceives plugins/*.py (skips init.py), scores plugins with drives (curiosity, mastery, stability, task-drive) and stages.
• Mutates code via patterns; tests with pytest; rolls back on failure.
• Tracks tasks from tasks/*.yml (phase/difficulty/category), per-task streaks in manager/tasks_state.json; advances phases when enough tasks are mastered.
• Tracks error streaks; on repeated failures (stage ≥2) consults whitelisted docs and graph API; stores text in external knowledge.
• Logs per-step diary with selections, actions, tasks, curriculum, reflections (via LLaMA), lifecycle events, meta-skill.
• Metrics recorded in manager/metrics.json.

State files:

• manager/brain_memory.json (observations, patterns, external knowledge, error/meta-skill).
• manager/concept_graph.json (plugin/function stats).
• manager/curriculum_state.json (phase/mastery).
• manager/tasks_state.json (task streaks/status).
• manager/mind_diary.json (per-step log).
• manager/metrics.json (steps, accepted/rejected, consults, tasks mastered).

Reset (fresh life):

python3 scripts/reset_state.py

Clears state JSONs and recreates sample plugin/task.

LLaMA/reflection config:

• MIND_LLM_BASE_URL (default http://127.0.0.1:11434/v1/chat/completions)
• MIND_LLM_MODEL (default llama3.2:1b)

Web/graph consults:

• Domains: docs.python.org, realpython.com, developer.mozilla.org.
• Graph API: GRAPH_API_URL (default http://127.0.0.1:8000).

Task format (tasks/*.yml):

name: list_sum
phase: 1
difficulty: 1
category: math
target_plugin: sample_plugin.py
target_function: process
description: ...
requirements:
  - process([1,2,3]) == 6
  - process([]) == 0
  - process([-1,1]) == 0

License: see LICENSE (all rights reserved by George Toloraia).

Architecture additions

• NeuronGraph (manager/neuron_graph.py): embedding-backed nodes for plugins/tasks/errors/reflections with weighted edges and Hebbian updates; persisted to manager/neuron_graph.json.
• Embeddings + Value Function (manager/embeddings.py, manager/value_function.py): pluggable embeddings with offline fallback and reward-informed scoring (USE_VALUE_FUNCTION=1).
• Multi-agent loop (manager/agents/): planner → coder → critic → memory agents orchestrated inside the life loop when HUMAN_MULTI_AGENT=1.
• Online learning + reward (manager/reward.py, metrics extensions): moving averages per plugin/strategy plus neuron-graph edge updates driven by rewards.
• Perception & practice envs (manager/envs/): structured error/percept logging and lightweight text/symbol environments (HUMAN_ENABLE_ENVS=1, HUMAN_ENV_RATIO).
• Reward shaping v2: structured outcome taxonomy (domain/tests/env progress) with clipped rewards and breakdowns stored in metrics history.
• Trace/debug + async scaffolding: set HUMAN_TRACE=1 to emit per-step traces (candidate scores, actions, rewards) under manager/traces/; set HUMAN_ASYNC=1 to use the async orchestrator wrapper for perception/env steps.
• Env curriculum scheduling: HUMAN_ENV_RATIO controls how often env steps run alongside code steps; domain choices are recorded in metrics for analysis and visualization (see export_for_viz in the neuron graph).

Reset and rerun

• Reset state (backs up JSONs to backups/<timestamp>/ and removes them):

  python3 scripts/reset_state.py
  

• Fresh run with tracing and value function, code-only:

  HUMAN_MULTI_AGENT=1 HUMAN_TRACE=1 USE_VALUE_FUNCTION=1 HUMAN_ASYNC=0 HUMAN_ENV_RATIO=0.0 python3 -m manager.life_loop
  

• Analyze the latest run:

  python3 scripts/analyze_brain_run.py