AI agents are getting better at retrieving context, memories and documents.
But retrieval alone does not answer three critical questions:
- Is this evidence authorized for this user or tenant?
- Does it actually support the final claim?
- Should the system abstain when support is insufficient?
That is the problem I’m working on with AletheionAGI.
The basic flow is:
retrieval / memory → authorized evidence → grounding → reader LLM → claim validation → deliver or fail closed
Aletheion is designed to sit on top of existing infrastructure rather than replace it. Teams can keep their current LLM, RAG pipeline, vector database and memory system.
So far, under controlled testing, we have validated:
- tenant isolation with identical identifiers across organizations;
- 200 hosted queries with no foreign canaries observed under the tested conditions;
- persistent memory with authorized retrieval;
- deterministic state reconstruction;
- API keys scoped by environment and namespace;
- BYOK reader inference;
- fail-closed behavior when evidence is insufficient.
We are now opening a small controlled cohort for proofs of concept and guided integrations.
The most interesting question for me is not whether an agent can retrieve the right information.
It is:
Can we make sure it refuses to use the wrong information, even when that information is available?
Project:
https://www.aletheionagi.com
Top comments (0)