MAXIM
A Bio-Inspired Cognitive Architecture for LLM Agents and Robots
What is Maxim?
Maxim is an open-source, bio-inspired cognitive architecture for LLM agents and robots, designed and built by Denny Schaedig and published on PyPI as pymaxim (Apache-2.0; current release 1.3.1, September 2026). It models brain systems — episodic and semantic memory, reward learning, fear, attention, proprioception — as a persistent substrate around the LLM, and tests it on a Reachy Mini robot and in a live Minecraft world it does not control.
Its claims are tested rather than asserted: every release since 1.0 has shipped with pre-registered, ablation-controlled experiments, and results are tracked in a public evidence ledger with the conditions that would require re-running them. The most recent: in 1.2 “Oasis” a want taught to one agent transferred to an independent agent over a real network, and in 1.3 a survival fear one agent learned from in-game pain transferred to an agent that never felt it — both with no LLM in the action path. 1.3.1 “Hardening” adds no new claim: it fixes what a re-score of 1.3.0 found, each fix shipped with the test or check that holds it.
These pages are the design story — why each piece is built the way it is, what the experiments showed, and what a real robot learned. The software itself is documented at pymaxim.bio: install, CLI, configuration, and the experiment ledger that decides what Maxim is allowed to claim.
The Problem with Robot Brains
Industrial robots execute precise, pre-programmed movements but adapt poorly when conditions change. LLM-driven agents can reason in language, but they typically carry no persistent model of their own body state (proprioception) and no learned memory of what worked in earlier sessions.
Maxim's approach is to borrow from neuroscience: rather than invent new mechanisms for memory, reward learning, and attention, it adapts computational models of the brain systems that handle those problems in animals — and then tests whether each one earns its place.
The Biological Blueprint
Maxim implements computational models of several brain structures, each handling a specific cognitive function:
Hippocampus
Episodic memory. What happened, when, and what came of it.
SCN
Circadian rhythms. Temporal patterns across hours, days, weeks.
Nucleus Accumbens
Reward prediction. Learning what leads to good (or bad) outcomes.
Entorhinal Cortex
Similarity matching. Finding relevant memories for new situations.
Pain Circuits
Nociception. Detecting and learning from aversive experiences.
Attention Networks
Focus. Deciding where to look and what matters.
IPS & Angular Gyrus
Dual number system. Fast approximate and slow precise math cognition.
Anterior Temporal Lobe
Semantic memory. Lasting knowledge distilled from experience.
Communication Gateway
External messaging. Reaching users via SMS, voice, and more.
Imagination
Real-time entity design from novel percepts during low-arousal states.
Default Embodiment
Base humanoid body with 5 sensors, 8 affordances, and 3 failure modes.
Operating Modes
Maxim's behavior is controlled by two independent dimensions: a processing state (awake or asleep) and an autonomy level (planning, supervised, or autonomous), each level capping how much initiative the agent may take without a human in the loop. An earlier layer of named strategies (observe, explore, research, assist…) was removed in the 1.0 mode refactor—it was policy written by hand, and the point of the substrate is to learn that.
Sleep isn't a mode. The agent enters it by calling the sleep tool, keeps its autonomy level, and wakes when input arrives. Like biological sleep it's active maintenance: memory consolidation and pattern extraction run while the lights are off.
How It All Connects
No single component carries the design; the point is how they integrate. Perception flows into memory. Memory informs attention. Attention shapes what gets remembered. Reward signals update causal models. Pain teaches avoidance.
An Illustrative Example
A robot is asked to "find the coffee mug." In the design, the systems are meant to cooperate like this:
- The Entorhinal Cortex retrieves similar past experiences (maybe "finding cups" or "kitchen tasks")
- The Hippocampus provides context: "Last time, the mug was on the counter"
- The SCN notes it's morning, when mugs are often near the coffee maker
- The Attention Network prioritizes scanning kitchen areas
- The Nucleus Accumbens predicts: "Counter search usually succeeds"
- If the robot moves too fast and jerks, the proprioceptive nociception signal registers an aversive event that the reward system learns to avoid
- Success or failure updates all relevant memories and predictions
Frequently Asked Questions
What is Maxim?
Maxim is an open-source, bio-inspired cognitive architecture for LLM agents and robots, designed and built by Denny Schaedig and published on PyPI as pymaxim. It implements computational models of brain structures (hippocampus, entorhinal cortex, nucleus accumbens, attention networks, proprioception, and more) as a persistent substrate that embodied agents use to perceive, remember, and decide.
What is the Maxim harness?
The Maxim harness is the orchestration layer that wires the biologically-inspired components (memory systems, attention, proprioception, reward prediction, tool use) together with an LLM and an embodied platform like Reachy Mini. It coordinates perception, memory recall, attention, and action in a single agent loop.
Is Maxim open source?
Yes — Maxim is open source on GitHub at github.com/dennys246/Maxim. The repository contains the full cognitive architecture, agent harness, simulation environment, and integration code for the Reachy Mini robot.
What hardware does Maxim run on?
Maxim runs on the Reachy Mini robot, processing real sensor input, motor commands, and human interactions. The architecture is designed to be hardware-agnostic via its embodiment layer, where a new platform declares its sensors and affordances; so far it has been validated on Reachy Mini.
How is Maxim different from other AI agents?
Most LLM agents keep no learned state between sessions. Maxim adds a persistent bio-substrate — episodic memory (hippocampus), semantic memory (ATL), reward prediction (nucleus accumbens), circadian rhythms (SCN), and proprioception — whose learned state, such as memories and reward associations, carries across sessions. Whether that state changes behavior is tested rather than assumed: the pre-registered 1.0 benchmark found that a strong LLM's prior usually dominates carried experience, while on a Reachy Mini robot a reward substrate with no LLM in the action path learned to turn toward speech in about ten trials and was correct at the start of the next session.
Design Goals — and What Has Been Tested
Biology is the design reference because it points at properties worth aiming for. These are design goals and working hypotheses, not demonstrated results; the evidence ledger tracks which have been tested and how:
- Robustness: brains operate in noisy, unpredictable environments
- Efficiency: biological systems work under tight energy budgets
- Continual learning: biological memory integrates new experience without wholesale overwriting of the old
- Embodied grounding: cognition tied to sensorimotor experience
- Safety: pain and fear are long-conserved harm-avoidance signals
The results so far are mixed, and the essays report both halves. On a Reachy Mini robot, a reward substrate with no LLM in the action path learned to turn toward speech — direction and step size — in about ten trials, and was correct at trial 0 of the next session from its persisted state (Sound Orienting). In simulation, the pre-registered 1.0 benchmark found that carried cross-session experience did not override a strong LLM's prior, so Maxim does not claim its LLM agents “learn and get better” (The Honest Benchmark). Since then the claims have moved into a world Maxim does not control: on a live Minecraft server, a taught want transferred between independent agents over a real network (1.2, n = 50 receivers per arm), and a fear learned from drowning pain transferred to 12 of 12 receivers against 0 of 24 isolated controls (1.3). The same releases publish their limits: the shared want is keyed to one exact situation and does not yet generalize across layouts, pooling learners helps each agent but costs more total experience, and a “dark means danger” fear was blocked because the sensors could not tell the situations apart (release notes).