Memory
Counted relationships, paired retrieval and sessions that can be saved and reopened.
Inspectable AI research · active development
UnisonAI asks whether memory, context, generation, perception and learning can be built around exact, inspectable records instead of native trained weights. The project stores what it has observed, counts the relationships used for a choice and records how each implemented component entered the system. It is active research: working mechanisms and open objectives are both kept visible.
Memory lab
Edit the example text, choose a word and inspect every continuation the small demonstration has actually observed. The result links directly back to those counts.
The architecture
Unison is not presented as a transformer with its weights removed. Its working design puts recorded memory at the centre, then builds selection, context, perception and action around information a reviewer can inspect.
Counted relationships, paired retrieval and sessions that can be saved and reopened.
A declared choice process over recorded candidates rather than an unexplained score.
Position-aware context, binding and checks against simply copying a stored sentence.
Image and audio experiments, voice, tools, observation and diagnostic components.
A generated token has a trace
Design history is part of the data
The implementation keeps a provenance label for each promoted component. That separates mechanisms directly inherited from the research model, necessary computational extensions, rebuilt established methods and current experiments.
Verification record
A registered decode instrument was also pointed at released neural networks, locating the placement law in specific projections and preserving its matched controls, ablations, and receipts.
The current campaign forward-forces a stronger generative surface, derives steering quantities from the law, and expands independent judged evaluation with every receipt preserved.
Reproduce it
python3 omni/benchmarks/verify_omni.pyRead the work
The repository documents UnisonAI's architecture, named verification campaigns, and the separate Fold Decode work on trained networks. The current V3 computation publication provides the model-wide computational foundation.