Preamble
This document extends the original noizu-labs-ml/artificial_intelligence library README (April 2018) which proposed that current-generation deep learning models omit critical aspects of biological neural architecture — dynamic firing thresholds, temporal decay, bidirectional feedback, and modular reusability. Concepts that have since been independently validated by Liquid Time-Constant Networks (Hasani et al., 2020), Transformer self-attention (Vaswani et al., 2017), and LoRA (Hu et al., 2021).
Rather than waiting for the compute and tooling to build this from raw neural primitives, this architecture takes a pragmatic approach: use LLM agents as stand-ins for biological lobes, with YAML-structured pipes as stand-ins for axonal connections. This lets us rapidly experiment with the behavior and auditability of a distributed cognitive system — observing how lobes interact, how shims shape perception, how memory consolidation affects long-running tasks — while keeping every message human-readable and debuggable.
The key insight is that the topology is the architecture, not the internals. GPT/Claude agents can be replaced over time with smaller fine-tuned models, then with compact trained neural modules. YAML interfaces can be replaced with raw vector embeddings — and critically, a bridging network can be trained to produce the same YAML output from those vectors, allowing old and new components to coexist during migration. Each replacement is a local swap, not a system redesign.
If initial outcomes from the Phase 1 prototype validate the behavioral model, the intended direction is embodied AI — where the time arrow (per-lobe temporal decay), autonomous processing units (lobes running independent tick loops), and global neuro signals (system-wide priority modulation) become essential rather than theoretical. A robotic system that needs to react to a snake on the ground in real-time benefits directly from lobes that maintain temporal continuity, shims that reshape perception based on environmental context, and a subconscious that accumulates slow-burn signals while the executive focuses elsewhere.
Executive Summary
This architecture models cognition as a distributed network of specialized processing lobes, each operating within a small, fixed-format context window. No single lobe holds the full state of the system. Instead, lobes communicate through interstitial routing nodes that apply contextual shims — active transforms that reshape data based on global neuro signals and upstream context before it reaches the next processor.
By constraining each lobe's input and output to a fixed contract and retaining per-lobe recency of prior values, the system processes tasks far larger than any individual context window could contain. A novel, a codebase, or a complex analysis is never “held” in memory — it is progressively digested through short-term memory, consolidated by a dream system into long-term storage, and recalled on demand through targeted queries.
Global neuro signals act as priority/urgency flags (not semantic content). They broadcast system-wide behavioral modifiers — analogous to neurochemicals — that each shim interprets differently. A “threat” signal causes the visual shim to boost snake-like pattern confidence while the analytical shim shifts to worst-case reasoning. Same flag, different behavioral responses per lobe.
Each lobe encodes a temporal arrow internally through its own sliding context window. The current tick's input appears at full resolution, while previous outputs and inputs decay progressively — the prior tick is lightly summarized, two ticks back is compressed further, three ticks back is a one-liner. This decaying trail of self-history biases each lobe toward continuity: a lobe that has been outputting “snake” for several ticks sees that conviction reinforced by its own recent context, without any explicit flag forcing it. As those echoes compress and fade, the conviction weakens naturally — the soft analog of neural firing threshold decay. This is the mechanism that makes a lobe “stay in a mode” and then gradually relax out of it.
The topology is fixed but evolvable. Phase 1 uses LLM agents passing YAML over MCP pipes. Phase 2 replaces YAML with raw token embeddings. Phase 3 replaces shims with LoRA adapters and lobes with compact trained modules. The wiring never changes — only what flows through it.
System Topology
Full architecture showing lobes, interstitial routing nodes with shims, memory tiers, and global neuro signals. Lobes never connect directly — all data flows through interstitial nodes that apply contextual transforms.
Interstitial Node & Shim Detail
Anatomy of an interstitial routing node. Data enters, passes through an input shim (contextual transform), hits the routing logic, then passes through an output shim before forwarding to downstream targets. Neuro signals feed into both shims.
Time Arrow: Per-Lobe Decaying Context Window
Each lobe maintains a sliding context window that encodes temporal state. Current input arrives at full resolution. Previous ticks' inputs and outputs decay progressively. This creates soft mode persistence: a lobe seeing its own recent “snake” outputs in context is biased toward continuing that assessment, but the bias fades naturally as those entries compress away.
Memory Tiers & Dream Consolidation
The three-tier memory system. Processing lobes write to STM through the Memory Gateway shim. The Dream System periodically scans STM, extracts entities and summaries, commits them to LTM, and prunes STM. Recall flows back through the gateway with relevance filtering.
Evolutionary Migration Path
The same topology across three phases. Internals change, wiring stays fixed.