Continual learning without full retraining
Project Nidra
A training-free continual-learning direction using episodic memory, routing, Oja-style consolidation, and bounded-memory evaluation to study whether models can keep learning without full retraining.
Core hypothesis
Cheap agentic intelligence needs memory that survives long streams. Nidra tests whether a frozen model can absorb new facts through external memory, sleep-style consolidation, and routing without catastrophic forgetting.
Steps taken
- Implemented episodic memory and NidraLM-style retrieval/logit-bias paths for fast factual updates.
- Added sleep consolidation with clustering, ZCA whitening, sparse Oja memory, decoder/prototype routing, and micro-sleep evaluation hooks.
- Built sequential benchmark runners for CounterFact, QAEdit, ZsRE/NCL-style hidden tests, streaming curves, BWT/FWT, conflict handling, compositional reasoning, and bounded-memory behavior.
- Explored a Nidra meta-learning branch where internal conflict changes later arbitration without labels or reward.
Results
Benchmarks are reported under a WILD-style evaluation frame: Without Intervention, Live Decoding, following The Mirage of Model Editing: Revisiting Evaluation in the Wild.
Llama-3-8B CounterFact: 79.6% reliability, 80.7% generalization, 32.3% locality.
QAEdit: 80.3% reliability, 79.3% generalization, 32.3% locality; Qwen2.5-7B CounterFact: 73.9% reliability, 75.0% generalization, 34.8% locality.
Qwen CounterFact 500-step stream ended around 0.7778 reliability/generalization with BWT 0.0.
NCL hidden-test sleep-only: ACC 63.3%, BWT +21.43%, FWT +3.6%, compositional reasoning 83.33%, conflict resolution 37.5%.
Meta-learning PoC: 7/7 behavioral invariants passed; disagreement-only learning changed halt sensitivity by +7.13% without external labels.
Caveat
The current system is closer to a serious continual-learning harness than a solved memory architecture. The next real bar is bounded long-stream learning with stable conflict revision.
Status
Prototype and benchmark harness in progress; roadmap is explicit about memory budgets, schemas, conflict revision, and meta-policy learning.