IONS: A reasoning graph that stores claims, evidence, and reasoning paths outside the LLM

I’ve been experimenting with an open source alternative approach to AI memory and reasoning called IONS.

The basic idea is that instead of storing all knowledge inside model weights, knowledge is represented as a graph of evidence backed claims called Cognitive Building Blocks (CBBs).

Each CBB contains:
\-A claim
\-Supporting evidence
\-Confidence metadata
\-Provenance
\-Relationships to other claims

Relationships are typed:
\-supports
\-causes
\-contradicts
\-depends\_on
\-derived\_from

When a query is executed, the system traverses the graph and returns:
\-The answer
\-Supporting claims
\-Confidence scores
\-The reasoning path used to reach the conclusion

The goal is not to replace LLMs.

The goal is to make reasoning and knowledge inspectable rather than implicit.

Current questions I’m exploring:
\-How does this compare to GraphRAG?
\-Does explicit claim storage improve explainability?
\-Can confidence be computed from evidence quality instead of generated by the model?
\-Can knowledge be shared across independent nodes without retraining models?

Public node:
162.243.203.243:8000
Whitepaper:
[github.com/nomad505050/ions-genesis/docs/whitepaper.md](http://github.com/nomad505050/ions-genesis/docs/whitepaper.md)

I’d appreciate feedback from anyone working on GraphRAG, knowledge graphs, memory systems, agent memory, or explainable AI.