KONTINUUM Core

Pure-Python, neuro-inspired learning engine extracted from KONTINUUM. No Home Assistant dependency — usable from any Python project. Zero runtime dependencies (standard library only), Python 3.9+.

Part of the KONTINUUM family: kontinuum-core (this repo, HA-free Python package on PyPI) · ha-kontinuum (full HA Pro integration with UI) · ha-kontinuum-lite (slim HA integration, no UI) · kontinuum-AI-anomaly (anomaly / novelty monitor for agent action streams, built on this engine)

Installation

pip install kontinuum-core

Usage

from kontinuum_core import KontinuumEngine

engine = KontinuumEngine()
engine.register_entity("binary_sensor.motion_kitchen", ha_area="kitchen", domain="binary_sensor")

snap = engine.observe({"entity_id": "binary_sensor.motion_kitchen", "new_state": "on"})
print(snap.surprise, snap.anomaly)   # 0..1 surprise, bool anomaly flag

What it does

One observe() pipeline drives 26 brain-region modules: thalamic tokenization, hippocampal n-gram memory, predictive surprise with a robust (median + MAD) adaptive anomaly threshold, cerebellar reflexes, basal-ganglia habits, an attention (reticular) burst filter for noise, and sleep consolidation (replay / prune / dream-recombine / synaptic homeostasis during quiet spells).

Since 0.5.0 the set also models the missing slow signals — a lateral habenula (anti-reward: stop re-proposing rejected actions), a subthalamic nucleus ("hold your horses" under conflict), a learned circadian clock (suprachiasmatic), an interval-timing stopwatch for recurring cadences (e.g. "every few weeks"), and the cortisol / acetylcholine / serotonin modulators plus BDNF use-dependent protection. All are scalar EMAs or tiny bounded maps, start neutral, and add ~0 ms/event.

State is persistent and bounded: engine.to_dict() / from_dict() round-trip the full learned brain (with a schema_version guard), and the learned maps are capped — safe to run for years on a Raspberry Pi.

Every observe() returns an EngineSnapshot with the live observability signals — surprise (0–1 prediction error) and the adaptive anomaly flag — which the HA integrations surface directly as entities (sensor.kontinuum_surprise, binary_sensor.kontinuum_anomaly, …). The engine stands on its own; any LLM is strictly an optional layer on top.

Documentation

Full reference in docs/: MODULES.md (all 26 brain modules) · PIPELINE.md (the observe() flow, EngineSnapshot.extra fields, the reward loop, persistence).

LLM integration contract (kontinuum_core.llm)

The engine is the sub-symbolic brain; an LLM is the optional language / reasoning layer on top.

• build_llm_context(engine_or_brain) / render_llm_context(ctx) — export the state (anomaly signal, expected-next events, learning maturity) with explicit 0–1 scales so a model can reason over it reliably.
• extract_json(reply) / normalize_proposal(reply) — turn a model's (often sloppy: code-fenced, prose-wrapped, stringly-typed) reply into a strict, validated action proposal.

Day-1 priors (kontinuum_core.priors)

parse_home_prior(llm_reply) + seed_engine_from_prior(engine, prior) let an LLM describe the home at setup, so the engine starts already expecting the household routine instead of from a blank slate.

Built on top of core

The engine is designed for smart-home event streams, but it is a general sequence-learner. kontinuum-AI-anomaly (PyPI: ai-kontinuum-monitor) points it at an AI agent's action log instead and adds the monitor layer core deliberately leaves out — robust scoring, anomaly history, alerting (escalation / snooze), a dashboard, multi-agent cross-stream correlation, strategy presets and an LLM-feedback loop. Core stays untouched there; it is a thin, additive layer, which is what keeps this engine equally usable from the Home Assistant integrations and from an agent-monitoring context.

pip install ai-kontinuum-monitor   # pulls in kontinuum-core

Benchmark

python benchmarks/replay.py

A replay benchmark + concept-drift stress test (also runs as a CI quality gate): the surprise signal separates anomalies from routine at AUC ≈ 0.99 and re-adapts after a routine change.

License

AGPL-3.0 – see LICENSE file.