100 small distributed LLMs smarter than 1 monolith 100 small distributed LLMs smarter than 1 monolith
    Research project - Working PoC

    AI you can extend, retrain, and forget on purpose.

    SmenaAI is a working modular AI architecture built from specialist nodes. Add a new domain by attaching an adapter. Retrain the part that needs it. Drop knowledge that has gone stale, or that should never have been learned in the first place. The rest of the system keeps working unchanged.

    Why study a different architecture.

    Consensus hides risk

    Polished answers flatten objections that should still shape the final decision. Once dissent disappears into a smooth summary, the system can sound aligned while dropping the exact risk signal a human needed.

    Durable expertise is hard

    Adding context helps temporarily. Durable expertise is still fragile. A system updated for one domain may quietly lose what it previously knew and not surface that loss when it matters.

    Broad models flatten in narrow domains

    General systems sound confident across many topics. They often stay shallow where legal, compliance, security, medical, and technical diligence demand depth.

    What the PoC is testing.

    The core bet is that domain depth should come from adding and adapting the right specialists, not from repeatedly rewriting one shared model. The PoC focuses on routing, bounded participation, disagreement, and local adaptation.

    Selective Activation

    Only the relevant foxlings and clusters should activate for a query, instead of waking the full system every time.

    Specialist Growth

    Rapid domain expansion should come from adding the right specialists, not from layering more generic critique onto the same broad model.

    Built-In Error Detection

    Critique has an explicit role in how answers are formed. Objections surface before synthesis, not after, which matters in high-stakes work.

    Adaptive Learning

    RAG changes what the model sees. Local adaptation changes how a node behaves. Smena tests when retrieval stops being enough.

    Why the architecture is different.

    The goal is to extend a reasoning system into new domains while preserving visible disagreement, local adaptation, and inspectable synthesis.

    What modular actually means in practice.

    Most AI systems are black boxes you can only grow by retraining the whole thing. SmenaAI is built to be edited. Four operations matter.

    Add a domain

    Drop a new specialist into the system with its own adapter, memory, and tools. The other clusters do not need to know it exists until the router decides they should.

    Retrain a part

    Update one cluster or one family of foxlings. Local retraining keeps the rest of the system intact and avoids the regressions that come with rewriting one shared parameter mass.

    Forget on purpose

    Pull a domain out without poisoning what stays. Knowledge that has gone stale, that was learned under bad data, or that should never have been there in the first place can be removed at the node level.

    Recompose the system

    The same specialists can be re-routed, regrouped, and reassembled for new tasks. The architecture is a pool of editable parts.

    Current PoC Snapshot.

    8

    cognitive clusters, each with a distinct reasoning role

    24

    foxlings in the core cognitive pool

    2-9

    foxlings activated per query - selective, not full-system

    3

    adaptation depths: memory, LoRA, fine-tuning

    Current Status.

    SmenaAI is a research program with a working PoC. The current focus is Stage A: testing whether routing specialization and disciplined memory beat simpler baselines on retention, stability, and cost.

    What exists now

    A working PoC with routed execution, cognitive clusters, bounded activation, inspectable debate traces, early dataset evaluation workflows, and support for custom foxlings with domain-specific adapters and tooling.

    What is being tested

    Retention under sequential multi-domain learning, routing specialization versus simpler baselines, and whether disciplined memory beats naive accumulation.

    What is not being claimed

    This is not AGI, not a finished enterprise product, and not proof yet that local adaptation beats every monolithic baseline.