---
title: Proto Cognitive Architecture
emoji: 🧠
colorFrom: purple
colorTo: blue
sdk: gradio
sdk_version: 5.31.0
app_file: gradio_app.py
pinned: false
license: mit
short_description: 'Neural Field + Episodic Memory + Hebbian Learning + Replay'
---

# Proto-Cognitive Architecture v5

A hybrid AI system combining a **Hebbian neural attractor field**, a **pinned episodic memory store**, and **Qwen2.5-0.5B-Instruct** — exploring how continuous neural dynamics can augment frozen language models.

## What's new in v5

**v4 proved** (via Suite 9 ablation) that the Hebbian field alone cannot recall facts — it scores 0.000 on all recall benchmarks. The field's real value is as a **familiarity detector** that routes queries via the CognitiveRouter.

**v5 makes the field actually learn structure:**

| Mechanism | What it does | Why it matters |
|-----------|-------------|----------------|
| **Hebbian W_local learning** | Co-activation updates to the region-to-region connection matrix | Regions that fire together *wire together* — real associative structure forms |
| **Memory consolidation (S→M)** | Stable, protected patterns slowly transfer to long-term memory tensor M | Knowledge persists beyond decay — M influences field dynamics continuously |
| **Replay / dreaming** | Self-training loop without external input: noise → field dynamics → Hebbian reinforcement | Reinforces learned patterns, strengthens weak connections, consolidates after |
| **Competition** | Winner-take-most suppression of weak activations | Promotes specialization — regions either contribute or go quiet |

## Architecture

```
Query → Tokenize → Per-token field update → Think loop:
  ┌─────────────────────────────────────────────────┐
  │  field_step()        — W_local + M influence    │
  │  competition()       — suppress weak regions    │
  │  attend(x_embed)     — input projection + top-K │
  │  update(attn)        — state update + decay     │
  │  hebbian_update()    — learn W_local structure  │
  │  apply_attractors()  — pull toward stored states│
  │  ↻ repeat N steps                               │
  └─────────────────────────────────────────────────┘
  → if teach: consolidate(S → M) + auto-dream
  → CognitiveRouter: resonance × retrieval → route
  → Build prompt (system facts + soft prefix for CONFIDENT/CAUTIOUS)
  → Generate with Qwen2.5-0.5B-Instruct
```

## Route decisions

| | Retrieval HIGH | Retrieval LOW |
|---|---|---|
| **Resonance HIGH** | 🟢 CONFIDENT | 🟡 CAUTIOUS |
| **Resonance LOW** | 🟠 UNCERTAIN | 🔴 DEFER |

## Key findings

- The Hebbian field is **not a memory** — it's a **familiarity/routing signal**
- Episodic store (PinnedEpisodicStore) handles all reliable fact recall
- Field value: cognitive resonance measurement → query routing → confidence calibration
- v5: W_local now forms learned structure (was random in v4) — this is the field's real potential

## How to use

1. **Teach** facts using the teach panel or batch import
2. **Ask** questions — observe the route decision (🟢🟡🟠🔴)
3. **Dream** — run replay cycles to reinforce structure
4. **Watch** W_local connections form in the diagnostics panel

## Paper

This architecture is being prepared for arXiv submission. The honest negative result (FieldOnly ≈ 0.000) reframed into the CognitiveRouter is the core contribution.

## Credits

Built by [Chris4K](https://huggingface.co/Chris4K) at [ki-fusion-labs](https://ki-fusion-labs.de)
MCP Innovation Award 2025