NovaCue: LLM Preference Learning

About

# NovaCue — AI Preference Learning Extension

**BETA** | Works with ChatGPT, Claude, Gemini, and Grok

> **Stop repeating yourself.** Let AI remember what you want.

AI remembers you — but it can't forget. When you evolve, your AI is stuck in the past. NovaCue's preferences evolve WITH you — old habits decay, new patterns reinforce, contradictions resolve automatically. No manual memory management. Your AI grows as you grow.

---

## The Problem

**AI assistants have amnesia.** Every conversation starts from scratch.

- *"I'm so tired of telling ChatGPT 'use TypeScript with ESLint rules' EVERY. SINGLE. TIME."*
- *"Why can't Claude remember I prefer concise answers? I've told it 50 times."*
- *"ChatGPT's custom instructions are too broad. I want different preferences for coding vs. writing."*

16,000+ monthly searches for "ChatGPT remember preferences." 15M+ power users experiencing this daily.

ChatGPT's "Memory" is manual. Claude's "Projects" are static. Gemini has nothing. NovaCue's automatic preference learning with semantic embeddings is genuinely novel.

---

## The Solution

NovaCue learns your preferences automatically and injects them contextually into your prompts.

1. **Learns Automatically** — No setup required. Use AI naturally, NovaCue learns what you like.
2. **Context-Aware** — Different preferences for different tasks. Python preferences for coding, formal tone for writing.
3. **Adaptive** — Current session takes priority. Say "use blue" today, it overrides your usual "red" preference.
4. **Cross-Platform** — One preference system for ChatGPT, Claude, Gemini, Grok, and custom sites.

---

## Key Features

### Smart Preference Learning

- **Instant History Processing** — Learns from existing chat history on page load in under 30 seconds
- **Automatic Detection** — Recognizes patterns like coding languages, color choices, formatting styles
- **Multi-Segment Processing** — Handles complex multi-topic prompts by splitting into semantic segments and extracting preferences per topic
- **Keyphrase Extraction** — Uses semantic embeddings to find meaningful bigrams (e.g., "bidirectional search") instead of isolated words
- **Distinctive Keyphrases** — Selects what's unique to each segment beyond the type name
- **Type-Based Organization** — Preferences grouped by context (TikZ Diagrams, API Design, Python Code)
- **Semantic Deduplication** — Prevents near-duplicate preferences by checking embedding similarity
- **Spell Correction** — Learns vocabulary from LLM responses and corrects typos via fuzzy matching
- **Smart Filtering** — Ignores greetings and social niceties — only learns from actionable prompts
- **LLM-Directed Proposals** — Preferences framed as instructions to the AI ("For Python Code: apply type hints")

### Session-Aware Intelligence

- **Instant Preference Display** — Preferences load instantly without waiting for model initialization
- **Live Auto-Refresh** — Preference count updates automatically as new preferences are learned
- **Momentum Scoring** — Recent preferences get 5x weight boost
- **Cross-LLM Session Continuity** — Switch from ChatGPT to Claude mid-conversation and pick up where you left off via "Continue Session?" popup within a 30-minute freshness window
- **Conflict Resolution** — Current intent overrides historical patterns
- **Context Fingerprinting** — Preferences tagged with co-occurring concepts to distinguish projects (e.g., "azure|python|sql server" vs "aws|typescript|dynamodb")
- **Recency Boost** — When no distinguishing keywords are present, most recently used project wins

### Episodic Memory Graph

- **Multi-Probe Retrieval** — Splits long prompts into sentences, queries past episodes per sentence. A brief mention of "MongoDB" in a 200-sentence prompt still finds your MongoDB conversations from 6 months ago
- **Graph Walk** — Finds the most relevant past episode, walks edges to discover connected conversations (2-hop)
- **Rich Context Injection** — Pulls actual Q&A snippets from past episodes, not just summaries
- **Edge Creation** — Semantic similarity + token overlap between consecutive episodes forms a knowledge graph

### Enterprise-Ready

- **Team Preferences** — Share and enforce preferences across your organization
- **White-Label Support** — Custom branding for enterprise deployments
- **Detailed Audit Logging** — Every injection logged with original prompt, augmented prompt, which preferences were applied, and injection size
- **Effectiveness Dashboard** — Per-preference analytics: application count, effectiveness score, usage trend, decay status
- **Privacy-First** — All data stored locally, optional encrypted cloud sync
- **Custom Site Support** — Add any LLM platform via selector configuration

---

## Who Is This For?

**Power Coders** — Consistent code style across sessions. Language-specific defaults (Python → type hints, JS → ES6). Framework preferences (React → functional components). NovaCue remembers your stack per project — Azure/Python preferences don't bleed into your AWS/Node.js project.

**Content Creators** — Maintain tone across AI-generated content. Platform-specific formatting. Brand guideline enforcement. Professional LinkedIn voice stays separate from casual blog tone.

**Academics & Researchers** — Automatic citation format (APA, MLA, Chicago). Academic tone preservation. Subject-specific terminology. Statistics paper preferences stay separate from philosophy essay preferences.

**Design Teams** — Design system adherence. Style guide preferences. Tool-specific defaults. Different projects with different brand guidelines stay organized automatically.

---

## Use Cases

**Developer Workflow** — Without NovaCue: "Create a React component" → class component → "Use functional components with hooks" → regenerate → next day, repeat. With NovaCue: functional components with hooks automatically. Saved 2 minutes, right first time.

**Content Creation** — Without NovaCue: "Write a LinkedIn post" → casual tone → "Make it professional, add hashtags" → rewrite → next post, same corrections. With NovaCue: professional tone with hashtags automatically.

**Multi-Context Work** — Morning: "Create TikZ diagram" → auto red boxes. Evening: "Create API design" → auto REST + JSON. Different context, different preferences, zero manual effort.

**Multi-Project Work** — Project A: Azure/Python/SQL Server. Project B: AWS/Lambda/DynamoDB. Mention "azure" → Project A preferences surface. Mention "lambda" → Project B. Just say "deploy the API" → most recent project wins. No manual tagging needed.

---

## How to Use

1. **Use AI as normal** — No setup required.
2. **NovaCue learns automatically** — After a few interactions, preferences are detected.
3. **Review in Preference Manager** — See learned preferences with their source prompts.
4. **Lock important ones** — Make a preference permanent.
5. **Delete unwanted ones** — Remove incorrect preferences.

**Advanced:** Manual preferences, default always-on rules, source tracking ("From: ..."), type-based organization, cross-platform sync, audit logs with full injection history, effectiveness dashboard, custom LLM site configuration.

---

## Architecture

**Core Engine** — Automatic preference detection, fast history processing (~30s), multi-segment prompt processing, semantic keyphrase extraction, distinctive keyphrase selection, preference decay (0.99 rate), semantic dedup (>0.8 merge), context fingerprinting, non-actionable prompt filtering.

**Vector Embeddings** — MiniLM-L6-v2 (384 dimensions), ONNX/WebAssembly SIMD, embedding cache for batch operations, background service worker for content scripts.

**Dynamic Semantic Typing** — Auto-detect conversation type via embedding similarity, new type discovery via keyphrase extraction, type centroid reinforcement, 0.65 similarity threshold.

**Markov Chain** — Transition matrix for type sequences, 70% semantic + 30% Markov blending, predictive preference pre-loading.

**Episodic Memory Graph** — Per-interaction episode creation, multi-probe retrieval (per-sentence queries), graph edges (semantic + token overlap + temporal), 2-hop graph walk, rich Q&A context injection, 30-day retention.

**Prompt Augmentation** — Non-actionable filtering → Default preferences → Type-matched preferences with conflict resolution → Curated session context → Markov predictions → EMG graph-walk context. 500-token budget. LLM-directed framing.

**RLHF** — Accept/reject proposals, vector alignment, type cluster reinforcement, rejection history (7 days), similarity duplicate check.

**Spell Correction** — Common misspelling map, fuzzy Levenshtein matching (distance ≤ 2), dictionary learning from LLM responses, applied before all tokenization.

**Storage** — chrome.storage.local (primary), IndexedDB (backup), chrome.storage.sync (cross-device), optional file system access. Cache hydration on startup.

**Platforms** — ChatGPT, Claude, Gemini, Grok, plus custom sites via selector configuration.

**UI** — Popup, Preference Manager (CRUD + source tracking + live refresh), Memory Journal, Insights Dashboard, Audit Logs with effectiveness tracking, Default Preferences, Cross-LLM session popup, AI Arena.

---

## Privacy and Security

- All data stored locally — preferences, episodes, and embeddings never leave your browser.
- Embedding model runs locally via ONNX/WebAssembly — no external servers.
- PII Detection — automatically detects emails, SSNs, credit cards, API keys, and more.
- Optional encrypted cross-device sync.
- Full audit trail for every preference injection.

---

*Built by LeadingTorch LLC*