Readable context
Keep retrieved material human-readable when traceability, citations, compatibility, or direct inspection matter most.
NORA
ResearchA research direction for giving AI systems useful external knowledge in the form that best fits the model and the task.
NORA explores retrieval augmentation as a flexible knowledge layer rather than a one-size-fits-all prompt technique. The project is centered on supporting readable text context, compact neural memory, and hybrid use of both while remaining open to different model families.
StatusResearch
FocusKnowledge augmentation
Context formsText / Memory / Hybrid
Model scopeModel-independent
[ ABOUT NORA ]
Bring external knowledge closer to the model.
NORA investigates how retrieval can become a reusable MLBricks capability that works across conventional language models, state-oriented systems, and future memory-aware models.
Compact model-facing context
Explore neural memory representations for systems that can work with retrieved information without relying only on long blocks of raw text.
Use both when useful
Combine readable evidence with compact memory so an application does not have to discard the strengths of either approach.
[ RESEARCH GOALS ]
Flexible retrieval for different kinds of intelligence.
NORA is currently a research project. The emphasis is on exploring useful behavior, interoperability, inspectability, and efficiency before making broader performance claims.
Work with external knowledge
Help models answer from information supplied by documents, knowledge collections, and application-specific sources.
Support different model families
Keep the retrieval layer usable with standard language models as well as MLBricks state- and memory-oriented research systems.
Make retrieval inspectable
Keep room for interfaces that show which information was selected and how it contributed to the final context.
Explore lighter context delivery
Study whether compact memory forms can complement raw text when large retrieved contexts become expensive to process.
RESEARCH STATUSNORA is under active exploration.
Its text, memory, and hybrid directions are research paths, not claims that one form is universally better than another. MLBricks intends to evaluate them according to task quality, latency, memory use, transparency, and model compatibility.
[ NORA / RESEARCH ]
Knowledge should fit the model,
not force the model into one format.
NORA is MLBricks' research direction for retrieval augmentation that can remain readable, become model-facing memory, or combine both depending on what the application needs.