What our R&D has produced.
One framework is published so far. New entries appear here as they reach specification and verification.
Novel Cognitive Learning (NCL)
A machine learning algorithm for continuous agent self-improvement at the application layer — without modifying model weights. NCL maintains a plastic memory graph whose nodes and edges carry measurable strength and salience. Five biologically-grounded mechanisms drive learning: prediction-error salience, Hebbian plasticity (LTP/LTD), spreading activation for associative retrieval, Ebbinghaus forgetting, and sleep-phase consolidation. Surprise — not repetition — is what makes an agent learn.
- An 11-step cognitive cycle: predict, act, measure, encode, consolidate, forget
- A three-outcome taxonomy — success, corrected, failure — so agents learn from human intervention without failing first
- 11 equations and 19 constants, verified line-for-line against a standalone reference implementation
More frameworks will be listed here as they are published.
From research to product, in three stages.
Formal specification
Every mechanism is written down as exact mathematics — equations, constants, algorithms — before any code claims it works.
Reference implementation
The spec is verified against a standalone implementation. If the computed values don't match the equations, the spec or the code changes — not the claim.
Product integration
Only verified frameworks inform the Memkist platform — the same discipline as the rest of the system: grounded behaviour, explainable results, no theatre.
Research you can read, product you can use.
Questions about the frameworks or how they shape Memkist? Get in touch.