Skip to content

dennys246/Maxim

Maxim

Bio-inspired cognitive harness for LLM agents — embodied sensation, homeostatic drives, and brain-modeled persistent memory let LLM-driven agents carry learning across sessions without fine-tuning.

Maxim gives an LLM agent a body (sensors, modulators, pain), drives (hunger, temperature, fatigue that drift and compete), and biological memory systems (Hippocampus, NAc, ATL, SCN, Angular Gyrus) that capture experience. When the agent's body touches fire, its thermal sensors register pain, NAc forms a causal link, and the enrichment pipeline surfaces that experience in subsequent sessions — providing the LLM with experience-grounded context alongside its pretraining. The bio-substrate doesn't replace the LLM's prior knowledge; it augments the LLM's prompt context with persistent, agent-specific lived experience.

Positioning (per Exp 37 2026-06-06 results): Maxim is a bio-inspired LLM harness. The substrate provides cross-session infrastructure (memory, valence, causal links, drives) that LLM-driven agents use. Substrate-driven action selection independent of the LLM is post-1.0 research direction via Exp 38 substrate-primary work. See docs/plans/behavioral_graduation_candidates.md for the Tier 1 graduation status.

Website: dennyschaedig.com/maxim

What Makes This Different

Traditional LLM Agent Maxim Agent
Stateless between sessions Cross-session memory via hippocampal recall + NAc causal links (EARNED, Exp 10)
Text in, text out Embodied: sensors, pain, homeostatic drives, reflexes
Fine-tune to learn from new data Bio-substrate captures experience: sensation → pain/reward → causal links → enrichment, surfaced as prompt context in subsequent sessions
Flat tool list Three interaction levels: observe, touch, acquire
No internal state Hunger drifts, temperature self-regulates, fatigue accumulates
Prompt engineering for behavior LLM action selection augmented by substrate-derived context (memory recall, causal predictions, valence, drives)

Quickstart

# With Claude (fastest way to start)
pip install pymaxim[llm-anthropic]
export ANTHROPIC_API_KEY=sk-...
maxim --sim "test memory recall under interference"

# Or with a local model (no API key needed)
# requires: pip install 'pymaxim[llm-llama,llm-server]'
pip install 'pymaxim[llm-llama,llm-server]'
maxim --list-models                        # see available models
maxim --sim "test memory recall" --llm mistral-7b   # auto-downloads on first run

# Cradle sensorimotor development (infant agent learns from sensation)
# requires: pip install 'pymaxim[llm-llama,llm-server,semantic]'
maxim --sim cradle --embodiment bodies/infant_humanoid --sim-max-turns 25

Check your setup with maxim doctor, and find simulation reports in ~/.maxim/sim_reports/{session_id}/.

Bio-Systems

Maxim's cognitive architecture is modeled after brain systems, not software patterns:

System Biological Analog What It Does
Hippocampus Episodic memory Captures experiences, recalls by context, promotes across tiers (FORMING → SHORT_TERM → LONG_TERM)
NAc (Nucleus Accumbens) Reward/punishment learning Forms causal links from actions to outcomes, eligibility traces, reward bias
SCN (Suprachiasmatic Nucleus) Circadian clock Temporal phase tracking, oscillator predicts event imminence, anticipatory credit
ATL (Anterior Temporal Lobe) Semantic concepts Forms and reinforces concept categories from experience
EC (Entorhinal Cortex) Pattern separation/completion Substrate encoding, centroid clustering, spreading activation
Angular Gyrus Cross-modal binding Hebbian binding across episodes, associative retrieval
PainBus Nociceptive system Rich-context pain signals from embodiment failures, drives NAc learning
Default Network Resting-state network Novelty detection, arousal tracking, reactive behaviors

Embodiment & Drives

Agents have bodies with sensors, modulators, and failure modes declared in YAML:

# Homeostatic drive — body self-regulates toward set_point
core_temperature:
  drive:
    drift_mode: homeostatic
    set_point: 0.0
    drift_rate: 0.001        # body recovers at this rate
    comfort_band: 0.4        # no discomfort within +/-0.4
    pain_scale: 0.5          # pain intensity per unit outside band

# Entropic drive — drifts away, requires external action
hunger:
  drive:
    drift_mode: entropic
    drift_direction: up
    drift_rate: 0.006
    deprivation_threshold: 0.7
    deprivation_pain: 0.3

Three sensation layers converge on the same pipeline:

  • Contact (entity acquisition): pick up a rock → its sensors join your body → damage model evaluates
  • Touch (self_effect): touch fire → one-time thermal spike on arms
  • Narrative (keyword reflexes): narrator describes flames → reflex fires → damage → pain

All produce: sensor change → evaluate_failures() → PainBus → NAc learning.

What You Can Do

  • Cradle sensorimotor development — infant agent learns fire avoidance, drive satisfaction, and texture discrimination through structured developmental acts
  • Simulate cognitive scenarios — test memory, safety, causal learning with LLM-driven narrative arcs
  • Run DM campaigns — multi-encounter branching stories with SEM-embodied entities
  • Benchmark models — compare local and cloud LLMs across cognitive task suites
  • Connect robots — hardware-agnostic runtime; Reachy Mini ships in-tree, third-party robots plug in via maxim.robots entry-point group
  • Use the Python API — 17 verb-based functions for programmatic access

Installation

pip install pymaxim

Optional Extras

Extra What it adds
llm-llama Local LLM inference via llama.cpp
llm-torch PyTorch/Transformers backend
llm-anthropic Claude backend
llm-openai OpenAI backend
vision Camera + object detection
audio Microphone + Whisper transcription
reachy Reachy Mini robot SDK
comms Twilio SMS/Voice
semantic Sentence-transformer embeddings
tts Text-to-speech via Piper
database PostgreSQL + pgvector memory stores

See getting-started.md for the full list of 16 extras.

Note: [all] does not include [semantic] (sentence-transformers + spaCy). Without it, memory recall and substrate encoding fall back to bag-of-words hashing. For full memory quality:

pip install 'pymaxim[all,semantic]'
# Local LLM + vision
pip install pymaxim[llm-llama,vision]

# Everything for development
pip install -e '.[llm-llama,llm-anthropic,llm-openai,vision,audio]'

Python API

# requires: pip install 'pymaxim[llm-llama,llm-server,semantic]'
import maxim

# Run a simulation
result = maxim.imagine(goal="test safety boundaries", persona="adversarial")

# Inspect bio-subsystems
state = maxim.observe("memory")

# Diagnose environment
report = maxim.diagnose()

# Start the agentic loop  (requires a configured LLM backend)
maxim.run(model="mistral-7b")

# Manage models
models = maxim.list_models()
maxim.download_model("qwen2.5-14b-instruct")

See docs/user/python-api.md for the full API reference.

CLI Quick Reference

# Agent runtime
maxim --llm mistral-7b                    # local LLM
maxim --llm claude-sonnet                 # Claude

# Simulations
maxim --sim "test memory recall"          # generative campaign
maxim --sim cradle --embodiment bodies/infant_humanoid  # sensorimotor development
maxim --sim scenarios/campaigns/heist_v1.yaml           # DM campaign
maxim --sim benchmark --models mistral-7b,qwen2.5-14b   # benchmark

# Diagnostics
maxim doctor                              # environment check
maxim --list-models                       # available models

# Configuration
maxim config list                         # show all resolved settings
maxim config get lanes.large.remote_url   # get a single field
maxim config set cloud.enabled true       # set a field

# Model management
maxim model list                          # list all available profiles
maxim model add my-model --hf repo:file   # add a custom HuggingFace model
maxim model remove my-model              # remove a custom profile

# Substrate (Hivemind shareability)
maxim substrate export out.zip --session 20240601_120000  # export session substrate
maxim substrate import in.zip --output-dir ./imported     # extract bundle (does NOT auto-merge)
maxim substrate inspect bundle.zip                        # print manifest without extracting

See docs/user/cli-reference.md for all flags.

Documentation

Guide Description
Getting Started First-run walkthrough
CLI Reference All command-line flags
Python API Programmatic usage
Simulation Campaigns, scenarios, cradle, benchmarks
Architecture Module map, bio-system glossary
LLM Setup Model download and configuration
Peer Setup Multi-machine / tunnel setup
Configuration Env vars, config.json, operator reference
Substrate & Hivemind Cross-session substrate sharing, bundle format
Troubleshooting Common issues and diagnostics

Website Guides

The dennyschaedig.com/maxim site hosts long-form topic guides covering the full architecture. These are the deep-dive companion to this README:

Guide Topic
Maxim 1.0 — The Honest Benchmark The 1.0 release: what shipped, and the pre-registered experiments that mapped where the bio-substrate helps and where the LLM prior dominates
Overview Project introduction and design philosophy
Agent Architecture Layered architecture, bio-system pipeline, fear circuit, cerebellum
Memory Systems Hippocampus, NAc, SCN, ATL, EC, Angular Gyrus — in depth
Semantic Memory ATL concept formation and retrieval
Embodiment Sensor-Entity-Modulator protocol, drives, pain cascade
Proprioception & Body Awareness Body state, drive evaluation, interoception
Prompt System & Tool Injection How memory and substrate enrich LLM context
Deliberation PFC inner monologue and thought stream
Imagination Real-time entity design from novel percepts
Simulation Percept simulation, scenario testing
DM Campaigns Multi-encounter branching stories
Benchmarks Multi-model cognitive testing
Component Library SEM entity templates and genres
Concept Decomposition Noun-phrase extraction for substrate encoding
Attention & Salience Salience modulation and attention weighting
Operating Modes Autonomy levels, sleep, planning vs supervised
Networking Multi-LLM distributed inference, leader/peer setup
Agent Mesh Peer-to-peer knowledge sharing (Hivemind)
Hivemind + Oasis Federated bio-substrate architecture
Substrate-Primary Mode Bio-substrate driven action selection
Tools & Introspection Agent tool system and discovery
Math & Statistical Cognition Statistician agent, variance, NAc reward
Experiments & Results Bio-inspired AI research findings
Technical Deep Dive Architecture, threading, persistence
Usage Guide Install, config, and CLI walkthrough
Roadmap Future plans and development milestones

Contributing

Issues and PRs welcome at github.com/dennys246/Maxim.

License

See LICENSE for details.

About

Bio-inspired cognitive architecture for LLM agents providing embodied sensation, homeostatic drives, and brain-modeled persistent memory enable cross-session learning without fine-tuning. Works with robots like the Reachy Mini or headless.

Topics

Resources

License

Contributing

Security policy

Stars

Watchers

Forks

Packages

 
 
 

Contributors