Every query.
One deterministic
operating layer.
The Intelligence Kernel is what makes any LLM sovereign. It wraps every model call in memory, governance, and compounding context — regardless of which provider renders the response.
Six stages. Deterministic. Running on every request your organization makes.
You decide the model.
The kernel decides what it knows.
LongStrider doesn't pick your LLM. It wraps whichever one you choose in four layers of sovereign control — applied to every single call, automatically.
Provider Control
Model-Agnostic
OpenAI, Anthropic, Ollama, or any endpoint. The intelligence layer is fully independent of the model. Swap providers without touching the kernel.
Context Control
CIP-Assembled
Every LLM call receives a Contextual Intelligence Package — not raw history. Structured, ranked, and governed. The model sees exactly what the kernel decides it should see.
Behavioral Control
Runtime-Governed
The RuntimePolicy — written by the nightly engine — sets tone, challenge threshold, and depth for the next 24 hours. Persistent across every session and every model.
Memory Control
Compounding
What was learned in January shapes how the model responds in June. Automatically. The kernel routes compounded intelligence into every response without being instructed to.
01 — Intent Router
Before any retrieval runs,
the query is classified.
The Intent Router reads the incoming query and assigns a retrieval strategy before a single database call is made. Five intent types. Each selects a different retrieval profile and CIP assembly pattern.
Five intent classifications
Vector similarity is axis one.
The floor, not the ceiling.
Every other retrieval system optimizes one signal. The kernel scores five simultaneously — semantic, gravitational, structural, relational, and longitudinal — then ranks by composite weight.
Vector embedding — the floor, not the ceiling. Every query starts here and is supplemented by four higher-order signals.
Frequency, recency, emotional density, and outcome correlation. What actually mattered — not just what surfaced.
Information inside a Knowledge Cluster scores higher. Clusters compound. Isolated memories do not. Architecture enforces this.
Co-occurrence inference. Alias resolution. The live Relationship Graph — not keyword proximity.
Recent, longitudinal, and historical weighting applied simultaneously — in proportion to the query context.
19 queries. Parallel.
One structured package.
The Contextual Intelligence Package is assembled from 19 parallel queries — each contributing a distinct layer of intelligence. The architecture is designed for graceful degradation: every layer contributes independently, so no single slow source blocks the response.
19 parallel queries — all sources simultaneously
CIP — Assembled Package
Written nightly.
Applied to every response.
The nightly engine doesn't just consolidate memory — it writes a RuntimePolicy that governs how the kernel behaves for the next 24 hours. Tone, depth, challenge thresholds, evidence gates. Persisted in the substrate. Applied automatically.
RuntimePolicy — example fields
Written at 02:00. Active until next cycle.
challenge_thresholdFloat 0–1. Controls how aggressively the system surfaces contradictions. Written by governance pass. Applied immediately.
response_depthEnum: concise / standard / deep. Set per entity profile nightly. Adapts to observed preference patterns.
tone_calibrationFormal / direct / collaborative. Derived from Communication Calibration principle and behavioral history.
evidence_gateBoolean. When active: assertions require observed supporting evidence before being surfaced.
recency_bias_flagTriggered when temporal analysis detects anchoring to short-term data. Surfaces longitudinal pattern explicitly.
05 — Agent Integration
Agents that work.
Then write back.
Orbital agents don't just surface findings — they write them back to the memory substrate. Every agent task adds to the compounding intelligence. The system learns from what it does on your behalf, not just from what you say directly.
Agent writes are first-class citizens — gravity-weighted, entity-resolved, and visible to the five-axis scoring engine on the next retrieval.
Research Agent
Scheduled
↳ Topic summaries → Knowledge Cluster
Pattern Monitor
Threshold
↳ Behavioral signals → Entity Profile
Decision Tracker
Event-driven
↳ Decision records → Decision Cluster
Narrative Builder
Nightly
↳ Arc updates → Narrative Substrate