DC-GEM — Differentiable Core + Gradient-Free Episodic Memory

A hybrid Continual Learning System (CLS) architecture: a slow-learning differentiable transformer core augmented with a gradient-free episodic memory (GEM). Built on pre-trained models (Qwen3-0.6B), no training from scratch.

Key Results

Metric	Value
Fact-recall (memON)	1.00 (3 query variants)
Fact-recall (memOFF)	0.00 (natural, not a trivial gate)
Coding T1 (oracle/memON/memOFF)	1.00 / 1.00 / 0.00
Retrieval hybrid @ N=1000	0.915
PPL ratio (GEM, no Sleep)	1.00 (zero forgetting)
Semantic retrieval (bge vs core mean-pool)	0.84 vs 0.04 (chance 0.02)

Main Finding

GEM enables fact recall that the core cannot do alone. With pre-trained Qwen3-0.6B:

• memON retrieves the correct fact and generates the correct answer (1.00)
• memOFF has no access to facts and fails (0.00)
• Zero weight modification → zero catastrophic forgetting (PPL ratio = 1.00)

Sleep (Stability-Plasticity Dilemma)

Parametric consolidation (Sleep-to-LoRA) was tested and failed at 0.6B scale:

• Full Fine-Tuning: high plasticity (train=1.0) but zero stability (PPL ~10^10)
• LoRA + KL: full stability (PPL ~1.00) but insufficient plasticity (train=0.05)

This is direct empirical evidence that factual knowledge should be stored in non-parametric episodic memory (GEM) rather than parametric weights at small scales.

Evidence

• FINAL_REPORT.md — full closure report
• runs/evidence/py/GATES.md — all phase statuses
• runs/evidence/py/PREREG_v1.md — v1 thresholds
• runs/evidence/py/PREREG_v2.md — v2 thresholds (Sleep revised)
• runs/evidence/py/LEGACY.md — pre-PREREG results labeled

Reproducibility

# Environment
python -m venv .venv-rocm
source .venv-rocm/bin/activate
pip install -r requirements.lock

# Tests
python -m pytest  # 4 passed

# Fact-recall proof
HIP_VISIBLE_DEVICES=0 python examples/run_fact_recall_proof.py --n-facts 20

# Sleep consolidation (v3, LoRA)
HIP_VISIBLE_DEVICES=0 python examples/phase4_v3_lora.py

Architecture

dc_gem/
  config.py       # all hyperparameters, seeds
  core.py         # Qwen3-0.6B wrapper + hidden-state access
  memory.py       # EpisodicMemory + Episode/Store
  retrieval.py    # dense k=16 + lexical/BM25 + rerank
  embeddings.py   # bge-small encoder
  inject.py       # text + hidden channels
  sleep.py        # consolidation (experimental)
  lora.py         # LoRA adapter
  eval/           # producer, scorer, watchdog, jsonl

License

MIT