BIM — Biologically Inspired Model

BIM is a four-generation research programme and evolving, building artificial intelligence as a unified biological substrate rather than a frozen statistical predictor. The system learns continuously from interaction, updates its weights in real time, runs on a single CPU core under 1 GB of memory, and avoids catastrophic forgetting through sparse activation.

The working paper

BIM: A Biologically Grounded Substrate for Continuously Learning Artificial Brains
Working paper · June 2026 · UnikAI Lab

The paper covers the full architectural lineage — BIM 0 through BIM 3 — including the chimerical phase failure analysis, the unified substrate specification, and experimental results on attractor convergence and online plasticity.

Keywords: sparse distributed representations, predictive coding, Hebbian plasticity, hippocampal pattern completion, hierarchical temporal memory, continuous learning, biologically inspired computation.

Generations

BIM 0 (2026) — Graph-based association loops. Hebbian edge strengthening over a directed graph. Proved blank-slate learning is possible from interaction alone. Graph representation grew super-linearly; replaced in BIM 1.

BIM 1 (2026) — Sparse Distributed Representations. 16,384-dimensional SDRs with exactly 64 active bits (0.39% sparsity). Numba JIT-compiled kernels: 160 tokens per second on a single CPU core. 100% one-shot recall on 20-token sequences.

BIM 2 (2026) — Surprise-scaled plasticity. Jaccard distance between predicted and actual SDR gates the per-step learning rate. Perfect predictions skip the kernel call entirely. Added second temporal-pool layer and synaptic homeostasis.

BIM 3 (current) — Unified substrate. Five biological modules cooperating under local learning rules. No backpropagation. No external memory modules. No tokenizer.

BIM 4 (2026) — The Digital Species Prototype. Added an autonomic drive system to power a reinforcement learning agent. Included a simulated chemical Brainstem (Dopamine, Acetylcholine), an Actor-Critic Basal Ganglia for motor actions, and Sleep Consolidation via the Hippocampus.

Architecture — BIM 3

Sensory Codec. Deterministic byte-level codec. Each byte (0–255) maps to a fixed SDR via hash projection. Fully invertible, no learned parameters.

Predictive Cortex. Three-layer hierarchical predictive cortex with corrected top-down apical feedback. Each layer predicts the active columns of the layer below it. Mismatch drives Hebbian growth on prediction misses and sequential LTD on predicted-but-inactive cells.

Neuromodulator. Combines rolling prediction error with per-symbol habituation to prevent saturation on common characters. Modulates cortical plasticity globally.

Synaptic Homeostasis. Rescales and prunes saturated permanences. Synapses below the 0.40 threshold are deleted. Keeps the substrate functional over long training runs.

CA3 Hippocampus. Auto-associative attractor network binding high-surprise cortical states. Retrieval by pattern completion from partial cues — 100% retrieval rate on 50%-corrupted inputs.

Properties

  • CPU-Native Execution: Operates entirely on a single CPU core, keeping peak memory under 1 GB (traced at ~635 MB including the CA3 Hippocampus weight matrix).
  • Low Power Envelope: Runs within standard consumer CPU thermal power limits (under 20–30W), eliminating GPU dependencies at both training and inference stages.
  • Continuous Online Learning: Adapts dynamically in real-time from a single sequential streaming byte input.
  • Full Interpretability: Every active column, synapse connection, and attractor retrieval is fully inspectable.
  • High Interference Resistance: Demonstrates 0.0% accuracy degradation on previously learned natural language when sequentially trained on source code.

Chimerical phase — what failed and why

An intermediate BIM 3 system augmented the biological core with classical engineering components: byte-pair encoding tokenisation, an external episodic memory store, a symbolic concept graph, and sharded multi-process weight merging. It failed three evaluation gates after 15 hours of training on a 768K-token corpus.

Root causes:

  • Synaptic saturation. High-frequency tokens connected to every cell, destroying the SDR separation that prevents catastrophic forgetting.
  • Silent top-down apical projection. Bias clamped to [0, 0.3] against a threshold of 0.5 — the feedback pathway never fired.
  • Reversed winner-cell scoring. Outgoing synapse counts used instead of incoming, selecting the wrong cells as winners during sequence memory.

The deeper lesson: biological substrates do not compose cleanly with engineering bolt-ons. Unification, not augmentation, is required.

The Great Bifurcation

Following the success of BIM 4, the architecture possessed two proven, yet distinct capabilities: an incredibly efficient predictive text substrate and a robust digital organism framework. To honor both destinies, the active development split into two distinct tracks:

  • Applied Sequence Track: BILM — The Biologically Inspired Language Model. An applied model that focuses on sequence memory and mitigating catastrophic forgetting, optimized for processing sequential data.
  • The Long Game: BIO — The Biologically Inspired Organism. The realization of the ultimate UnikAI mission: a continuous-learning, autonomous artificial lifeform with simulated brain chemistry.
Blog: BIM hypothesis →All releases →