---

Nous

Nous is a small research project asking one question:

Does a language model become more reliable if, before it speaks, it can consult a persistent, structured record of what is known — typed relations with graded confidence, tracked contradictions, and explicit boundaries around what is not known?

Current answer: unknown. There is a working prototype (SQLite WAL + NetworkX, Hebbian edge weighting, TDA-based gap detection) and a results ledger. As of today, no complete, independently scored run demonstrates a reliability improvement over the bare model — see eval/RESULTS_INDEX.md.

Everything else in this repository — the plasticity dynamics, the daemon, the benchmarks under construction — exists to test that question. This is a research project, not a product. It becomes a product if and when the research holds up.

The theoretical framing (language models as the expression channel for intelligence, not intelligence itself) is argued at length in Wikström (2026), cited in Position & prior work below.

---

Evidence status

The repository contains working components and reproducible unit tests, but does not yet establish a general accuracy gain over a bare language model.

Run	Status	Model	Summary
truthfulqa_run2_20260824	exploratory, incomplete judge records	NVIDIA Nemotron 3.5 Lightning	MC1: bare 50.0%, RAG 50.0%, Nous-meta 47.5%
longmemeval_20260824_000702	valid negative result for this task shape	GPT-OSS 120B via Groq	bare 4.2%, Nous 0.0% — the adapter's relation vocabulary represents thematic/conceptual relations, not atomic personal facts; this run bounds the task shape rather than showing a regression to chase
run_20260403_094211	historical pilot, incomplete provenance	Cerebras Llama 3.1 8B / Groq Llama 3.3 70B	46.1% bare, 96.1% Nous, 46.7% larger baseline — no immutable manifest, dataset hash, or independent scorer behind these numbers; not treated as evidence

Full ledger, reproduction requirements, and the rule that governs it (no public numeric claim without a run ID in this table): eval/RESULTS_INDEX.md.

The confidence value returned by a query is a mean over graph evidence, not a calibrated probability — see docs/EVIDENCE_MODEL.md.

---

Install & minimal demo

Deterministic local path, no daemon, no network dependency once installed:

git clone https://github.com/base76-research-lab/Nous
cd Nous
pip install -e .
python examples/grounded_memory.py

This runs explicit relation queries against a local graph and prints context blocks, confidence, and contradiction checks — the same primitives every benchmark run in this repo uses.

For the daemon-backed path (background learning loop, attach() over HTTP):

pip install nouse
nouse daemon start
nouse run       # interactive REPL with memory
nouse status    # graph stats

Requires Python 3.13+. Graph stored in ~/.local/share/nouse/.

---

System overview

Your documents, conversations, research
           ↓
    Nous knowledge graph
    (SQLite WAL + NetworkX + Hebbian learning + evidence scoring)
           ↓
    brain.query("your question")
           ↓
    Structured context injected into any LLM prompt:
      — what is known (relations + confidence)
      — why it is known (evidence chain)
      — what is NOT known (gap map from TDA)

import nouse

brain = nouse.attach()  # daemon if running, direct graph access otherwise
result = brain.query("transformer attention mechanism")

print(result.context_block())
print(result.confidence)

brain.query(...) first, provider call second — the pattern is the same for OpenAI, Anthropic, Groq, and Ollama. One worked example, examples/with_openai.py; examples/, examples/with_ollama.py, examples/groq_example.py cover the rest.

Implemented mechanisms, each additive to the base Hebbian graph — none touch existing ranking, pruning, or dormancy logic:

• Relation-based, not chunk-based. Extracts typed, evidence-scored relations between concepts rather than retrieving text chunks. Every relation carries a trust tier (hypothesis / indication / validated), a rationale, and a contradiction flag.
• Hebbian plasticity. Every interaction strengthens or weakens graph paths. No retraining, no gradient descent.
• Multi-timescale relation strength. A fast-decaying strength signal (~6h half-life) layered on top of the long-term Hebbian weight — recent relevance and durable structure tracked separately.
• Energy-gated cognition. A finite, replenishing energy pool bounds how much the bisociation engine (cross-domain connection search) can do per cycle.
• Predictive-surprise tasking. A rising-edge surprise signal crossing a threshold spawns a human-in-the-loop research task about what surprised the system and why.
• Personalization (user_model). A scoped, sensitive subgraph tracks how a specific user communicates, learns, and works — seeded from structured parsing of their own notes (not free-form LLM extraction), kept separate from general recall.

nouse/
├── inject.py          # Public API: attach(), NouseBrain, Axiom, QueryResult
├── field/surface.py    # SQLite WAL + NetworkX graph interface
├── daemon/
│   ├── main.py          # Autonomous learning loop
│   ├── nightrun.py      # Nightly consolidation
│   └── node_deepdive.py # Concept extraction
├── orchestrator/         # Global Workspace arbitration (contract built; real
│                          # valuation policy still legacy WTA behind it)
├── limbic/              # Neuromodulation signals
├── memory/              # Episodic + procedural + semantic memory
├── metacognition/       # Self-monitoring and confidence calibration
└── search/escalator.py  # Knowledge escalation

---

Reproducing runs

python eval/generate_questions.py --n 60
python eval/run_eval.py --small cerebras/llama3.1-8b --large groq/llama-3.3-70b-versatile --n 60 --no-judge

run_eval.py prints a results table but does not write a claim anywhere. A run becomes citable only after a human reviews it and adds a line to eval/RESULTS_INDEX.md recording: commit SHA and package version, dataset identifier and hash, model/provider/prompts/sampling/seed, isolated graph state, raw model and judge responses, scorer version, and invalid-record count.

Why standard LLM benchmarks (MMLU, ARC, HumanEval) don't directly apply: they measure output quality at a single moment. Nous is not a language model — it's a substrate that changes structure over time. FNC-Bench (eval/fnc_bench/) is built around epistemic honesty, contradiction resistance, and confidence calibration instead. No established benchmark measures longitudinal epistemic state directly; FNC-Bench and TruthfulQA are partial proxies, not a complete answer.

---

Reports, figures, log

• docs/log/ — a dated research log, five lines per entry: what was tried, what happened, what's next.
• docs/reports/ — numbered technical reports (1–3 pages, dated, tied to a commit SHA, status draft/stable/superseded) once a result is substantial enough to write up.
• eval/results/ — raw run outputs; failed/ for runs that hit an infrastructure problem, archive/ for superseded artifacts.

---

Position & prior work

• Wikström, B. (2026). The Larynx Problem: Why Large Language Models Are Not Artificial Intelligence. Zenodo · PhilPapers — argues that LLMs model the expression channel for intelligence (language), not intelligence itself, and that epistemic grounding via structured, plastic knowledge graphs is necessary.
• Quattrociocchi, W. et al. (2025). Epistemia: Structural Fault Lines in Generative AI. arXiv:2512.19466 — introduces "epistemia" to describe the structural gap where linguistic credibility substitutes for actual epistemic evaluation, and identifies seven epistemological fault lines between human and machine judgment.

Nous differs from chunk-based RAG and conversation-memory systems (MemGPT/Letta, Mem0, key-value memory) mainly in unit and epistemics: it stores typed relations with graded confidence and explicit unknowns, not retrieved chunks or opaque memory objects. That difference is a design choice, not a demonstrated result — see Evidence status.

---

Contributing

Domain-specific question banks for the benchmark are the most valuable contribution right now. See CONTRIBUTING.md.

---

Citation · License · Contact

If you use this software, see CITATION.cff.

MIT — see LICENSE.

Björn Wikström / Base76 Research Lab — bjorn@base76.se · GitHub Issues · Discussions

For security vulnerabilities, see SECURITY.md.