Persistent learning
New facts survive restart without gradient updates.
Experimental AI Research · Alpha
PallattuLM
An experimental language model and continual-learning architecture exploring persistent memory, consolidation, pattern discovery, and sparse neural plasticity.
Stable core33.9MDecoder-only transformer trained from scratch.
Selective learning0.0061%Parameters touched per Lab #006 learning event.
Research question
Can a small stable model keep learning without repeatedly retraining the whole network?
PallattuLM-X tests whether new knowledge can first enter persistent memory, consolidate into reusable structure, and selectively modify tiny neural regions while the main language core remains stable.
Reproducible experiments
Six labs, one progression
These are controlled research results, not claims of general intelligence or production readiness.
Continual memory
Scaled memory, conflict updates, paraphrase recall, and abstention.
Consolidation
4,000 experiences compressed into semantic rules plus exceptions with 24.28x physical compression.
Pattern discovery
28 useful rules discovered from noisy observations with 100% held-out generalization in the controlled benchmark.
Selective neural plasticity
Sparse routed learning reached 98.8% of the full-control quality while touching less than 1% of parameters per learning event.
Transformer integration
The trained 33.9M PallattuLM core reached 100% on both held-out tasks while only 2,052 parameters, 0.0061%, changed per learning event.
Current status
Publishable research, experimental usability
Ready nowResearch
What is real today
PallattuLM has its own tokenizer, transformer, training and inference stack, persistent memory, consolidation, autonomous pattern discovery, and selective plasticity integrated with a trained transformer checkpoint.
- Base model
- 33,890,816 parameters
- Training corpus
- About 31M tokens
- Best selective result
- 100% controlled held-out accuracy with 0.0061% parameter updates per event
NextAlpha
What still needs to be proven
The model is not yet a production-ready general assistant. Natural-language continual learning, broad language quality, hallucination behavior, long-context robustness, safety, and real-world search-backed use still need evaluation.
Open research