Deep Dive
The NORDON Intelligence Engine
18 ML modules. Zero cloud dependencies. Built in Rust. Here's exactly how the engine works under the hood.
Architecture
The ML Pipeline
Ingestion Pipeline
Retrieval Pipeline
16 Modules
Every module, explained
Each module handles a specific aspect of the intelligence pipeline. All written in Rust. All run locally.
Importance Scoring
The core scoring engine. Scans every event for 40+ signals across 4 categories — architecture decisions, performance insights, security concerns, and dependency changes. A weighted algorithm rates each event from 0 to 1, with configurable boosts for user-flagged items (0.4), decisions (0.3), and failures (0.25). Only events scoring above the 0.4 threshold become memories.
40+ signal detectors, weighted scoring, configurable thresholds
Retrieval Ranking
A 6-signal ranking formula that finds the most relevant memories for any given context. Weighs keyword match (30%), importance score (20%), recency with 14-day half-life decay (15%), acceptance feedback (15%), scope match (10%), and frequency (10%). Handles token budget allocation to maximize context quality within LLM limits.
6-signal ranking, 14-day half-life decay, token budgeting
Pattern Detection
Detects recurring patterns in your codebase activity. Identifies file groups that are frequently modified together, failure loops where the same error recurs, and workflow patterns that indicate established team practices. Surfaces these patterns as memories so your AI can anticipate needs.
File group detection, failure loop detection, workflow patterns
Drift Detection
Monitors for architecture drift and stale memories. Detects when current coding activity contradicts stored decisions or patterns. Flags memories that haven't been accessed or validated in a configurable window. Sends drift alerts so your team stays aligned with documented decisions.
Architecture drift alerts, stale memory detection
Semantic Search
Hash-based embeddings for fast semantic similarity search with zero external dependencies. Generates compact vector representations of memory content and computes cosine similarity for retrieval. No API calls, no Python, no heavy ML frameworks — just pure Rust math.
Hash-based embeddings, cosine similarity, zero dependencies
Auto-Extraction
NLP-powered extraction that automatically identifies and structures important information from raw tool results and conversation context. Classifies extracted content into 7 memory types — decisions, failures, procedures, constraints, patterns, facts, and context snapshots — using pattern matching and heuristic analysis.
NLP extraction, 7-type classification, structured output
Dependency Tracking
An if-X-then-Y rule engine that tracks relationships between memories. When a decision depends on a constraint, or a procedure relies on a specific tool version, the dependency tracker maintains those links. Surfaces dependent memories together and warns when dependencies change.
If-X-then-Y rules, dependency graphs, change warnings
Memory Deduplication
Embedding-based duplicate detection that prevents memory bloat. Compares new memories against existing ones using semantic similarity, not just exact string matching. Merges near-duplicates intelligently, keeping the most complete and recent version while preserving unique details from each source.
Semantic dedup, intelligent merging, bloat prevention
Policy Engine
The security backbone. Scans all content for 31 distinct secret patterns (AWS keys, GitHub tokens, Stripe keys, JWTs, database strings, and more) and redacts them before storage. Blocks 28 sensitive file types from ever being read. Enforces team-defined retention and access policies.
31 secret patterns, 28 blocked file types, policy enforcement
Memory Decay
Four configurable decay profiles that control how memories age over time. Memories can decay linearly, exponentially, step-wise, or not at all. The decay function integrates with retrieval ranking so that aging memories naturally lose priority unless they continue to be accessed or validated.
4 decay profiles, configurable aging, access-aware
Feedback Learning
Closes the learning loop. Every time you approve or reject a surfaced memory, the feedback signal adjusts future scoring weights. Acceptance rates accumulate over time, making the importance scoring more accurate for your specific workflow. The system literally gets smarter the more you use it.
Acceptance tracking, weight adjustment, continuous learning
Confidence System
Memories get stronger or weaker over time based on usage. When a memory is accessed and leads to a good outcome, its confidence increases. Memories that go unused gradually lose confidence. This creates a natural selection process where the most useful knowledge stays prominent.
Confidence boost on use, decay on neglect, natural selection
Session Summaries
Automatically generates a structured summary when a coding session ends. Captures what was accomplished, what decisions were made, what failed, and what's left to do. The next session starts with full context of where you left off.
Auto-generated summaries, session continuity, context snapshots
Conversation Extraction
Extracts structured knowledge from natural conversation. When you discuss architecture, debug a bug, or explain a constraint to your AI, the conversation extractor identifies decisions, failures, constraints, and patterns from the dialogue and creates typed memories.
Chat parsing, decision extraction, constraint detection
Memory Compression
Token-efficient memory compression that reduces injection payload size without losing critical information. Intelligently summarizes verbose memories, strips boilerplate, and packs the maximum amount of context into the available token budget.
Token optimization, lossless compression, budget efficiency
Memory Quality Scoring
Grades every memory from A to F based on completeness, specificity, and actionability. Generates actionable suggestions for improving low-quality memories. Calculates an overall repo health score so you know the state of your project's knowledge base at a glance.
A-F grading, quality suggestions, repo health score
Performance
All ML runs locally — zero API calls
Engineering
Why Rust?
We chose Rust deliberately. Not because it's trendy, but because every alternative had a dealbreaker for local-first ML.
Memory safety without garbage collection
Rust's ownership model prevents memory leaks, use-after-free, and data races at compile time. No GC pauses means predictable performance — critical when you're scoring events in a hot path.
Cross-platform native binaries
A single codebase compiles to native binaries for macOS (ARM + Intel), Windows, and Linux. No JVM, no Python runtime, no Node.js. Just a binary that works.
Predictable performance
No JIT warmup, no garbage collection pauses, no interpreter overhead. The first event scores as fast as the millionth. Sub-5ms scoring is consistent, not a best-case number.
Single binary deployment
The entire ML engine ships as a single binary with zero runtime dependencies. No pip install, no npm install, no Docker required. Copy the binary, run it, done.
By the Numbers
The full codebase
Visualization
Knowledge Graph
The Knowledge Graph is built from memories and their relationships. Each memory becomes a node, and dependency links, co-occurrence patterns, and explicit references become edges. The result is an interactive force-directed visualization of your project's entire knowledge structure.
How the graph is built
Every memory creates a node. Dependency tracking (Module 07) creates edges between related memories. Pattern detection (Module 03) identifies clusters. Conversation extraction (Module 14) adds relationship metadata. The graph grows organically as your project evolves.
What you can do with it
Explore your project's knowledge visually. Filter by memory type to see only decisions, or only failures. Click a node to see its full content and connections. Identify knowledge gaps -- areas of your codebase with few memories. Share graph snapshots with your team.
Get Started
See the engine in action
Install NORDON and watch the ML engine learn from your first coding session. Free tier includes full access to all 16 modules.
npm install -g @sodasoft/nordon-cli