Who reasons
EVA runs steps under contract (local or remote LLM). It does not change catalogs, policies, or global memory alone.
Local-firstAuditableOn-premZero auto-merge
a cognitive kernel
Govern agents that learn from every run—not just orchestrate LLMs
EVA reasons. NOUS governs.
Every completed workflow or task feeds memory for the next: more efficient and grounded in real outcomes. While running, the kernel supervises reasoning and intervenes on sustained degradation—with explicit policies and audit, not model self-rule.
- Memory that accumulates Each run teaches the system; the next recalls verified experience.
- Live supervision The kernel acts on sustained confidence signals, not one-off noise.
- Explicit governance Policies, authorization, and auditable lineage—the LLM is not the admin.
The problem
Today’s AI agents are often a reactive wrapper around an LLM — insufficient for regulated production:
- No persistent memory beyond the context window.
- No learning across tasks: every conversation starts from zero.
- Weak governance: the model proposes tools without a kernel enforcing policy.
- No reproducible audit: reconstructing “why” means guessing from scattered logs.
The proposal
“EVA reasons. NOUS governs.”
EVA is the reasoning unit (LLM, e.g. llama.cpp). NOUS is the kernel: policies, memory, learning on graph and vector indexes, and auditable event sourcing.
How NOUS learns
Orchestrators connect steps between model calls. NOUS runs workflows and tasks in production: each completion consolidates experience; the next run reuses it—with a kernel that governs what is stored and allowed.
Per workflow or task
1 Workflow / task
Governed instance in runtime
2 EVA reasons
Steps, tools, and catalog skills
3 Kernel observes
Policies, live confidence, recall when needed
4 Memory
Episode + graph on close
5 Audit
Lineage; zero auto-merge
On close: episode + graph update. Auditable events in the EventStore.
Consolidation between runs
When the system is idle or between operational cycles:
ImprovementsRegressionsRepeated failuresStable successes
Governed replay Decay, pattern promotion, pause meta-policy
Workflow proposal Only if governance authorizes—never auto-merge
Long-horizon memory and policies evolve without the LLM rewriting the system.
- EVA — reasoning
- NOUS — governance & observation
- Persistent memory
- Authorization & audit
Unlike optimizing a text “skill” document (e.g. SkillOpt-style approaches), NOUS learns from the graph and episodes of each real execution under kernel governance.
How NOUS governs
The model proposes steps; the kernel decides what is allowed, persisted, and demonstrable.
EVA reasons. NOUS governs.
Who decides
NOUS enforces cost limits, collisions, and cancellation; supervises live confidence; injects context or escalates; persists memory on close.
Who operates
Control plane (nous_ctl) separate from product traffic: configure without mixing chat or gateway.
- Zero auto-merge: workflows, skills, and sensitive mutations require explicit authorization.
- Event sourcing: relevant decisions with lineage, not text logs alone.
- Skills and workflows: catalog invoked at runtime; learning lives in memory and graph.
Differentiators
Versus agent orchestrators (LangGraph, CrewAI, prompt chains) and optional RAG-style memory:
| Orchestrators / typical agents | NOUS |
|---|---|
| Chain LLM calls | Kernel governs full lifecycle |
| Every run starts at zero | Learns from each closed task |
| Optional vector store | Episodes + graph + multi-view recall |
| Model “in charge” | EVA reasons; NOUS authorizes & persists |
| Logs for debugging | Auditable event sourcing |
| Cloud-first | Local-first (llama.cpp + GGUF) |
Explore NOUS
Does this fit your case?
- Deploy a governed cognitive kernel in regulated production with traceability.
- Replace static RAG with online learning under explicit policies.
- Run cognition on-prem for compliance or latency.
- Contribute to an open-source kernel where design is the product.