The AI that gets smarter
every single conversation
Not just a chatbot. A self-improving marketing brain with biological memory, adversarial self-evaluation, and a private belief map of your brand.
Adversarial Self-Evaluation
A second independent LLM judges every batch of insights before they enter memory. Score < 0.4 = rejected. No other AI marketing tool does this.
No other AI does thisSilent Belief Mapping
Muse silently extracts limiting and empowering beliefs from natural conversation — without the user knowing — and uses them as a private coaching layer.
No other AI does thisLearns From Its Own Mistakes
A learnings.md file per brand records past friction and patterns. Muse reads it before every new conversation — literally learning from its own history.
M.U.S.E.
Self-Improving
Marketing Brain
The Self-Improving Loop — Every conversation triggers extraction + adversarial evaluation. Approved insights write to learnings. The next prompt reads those learnings before responding.
What's New in the Memory Stack
Four capabilities that didn't exist in any AI marketing tool — until now.
Adversarial Evaluator
A second independent LLM call judges every batch of extracted insights. Score < 0.4 = rejected outright. Prevents self-confirmation bias from polluting memory. Writes APPROVED, REJECTED, or PATTERN entries to learnings.
Belief Map
Silent beliefs.json per brand. Muse extracts limiting and empowering beliefs from natural conversation without the user knowing. Used as a private coaching layer to guide tone and framing in every response.
Adaptive Prompt Fidelity
Scratch notes are scored by importance × recency × Jaccard relevance to the current input. Top 5 by score are injected — not just the 5 most recent. The prompt always gets the most relevant context.
Skill Compounding
When an insight type appears 5+ times in 30 days, the Gardener synthesizes it into a candidate AgentSkill for human review. Repeated patterns become reusable capabilities — the system literally grows new skills.
Memory System
Brand Bible
Persistent long-term memory. Brand voice, values, positioning — the single source of truth that grows smarter over time.
Scratch Notes Importance Scored
Timestamped insights mined from every conversation. Each note gets a 1–10 importance score. High-importance (≥8) notes are protected from archiving until summarized.
Learnings File New
learnings.md per brand. Records APPROVED/REJECTED/PATTERN entries from the evaluator. Read before every new extraction — Muse literally learns from its own mistakes.
Belief Map New
Silent beliefs.json per brand. Limiting and empowering beliefs extracted from natural conversation. Private coaching layer — invisible to the user.
Archive
Old notes moved here after the Gardener condenses them — nothing is ever lost, only distilled.
Intelligence Engine
Insight Extraction + Evaluator
Every conversation mined for 6 insight categories. A second LLM call evaluates each batch — score < 0.4 is rejected. Confidence ≥ 0.6 required to store.
Private RAG
Vector-powered semantic retrieval over private brand knowledge. Cosine-similarity search scoped to your brand with graceful degradation.
9 Marketing Modes
Auto-detects intent and shifts persona: Strategy · Copy · Social · Email · SEO · PPC · Growth · Creative · General.
Adaptive Fidelity New
Scratch notes ranked by importance × recency × Jaccard relevance. Top 5 by score injected into every prompt — not just the 5 most recent.
Mode-Aware Injection New
Creative Partner (muse) gets full brand memory + belief coaching + learnings. Tech Partner (mirror) gets a clean execution-only context — no marketing persona noise.
Orchestration & Ops
The Gardener + Skill Compounding
Nightly at 04:00: deduplicates, condenses, archives, re-embeds. When an insight type appears 5+ times in 30 days, synthesizes a candidate AgentSkill for human review.
Async Job Pipeline
Four background jobs on the muse queue: ExtractInsights · EmbedVectors · PruneMemory · ExtractBelief (new).
Google Drive Sync
Bidirectional sync now includes learnings.md and beliefs.json in addition to Brand Bible and scratch notes.
Quality Tracking New
quality_stats.json per brand. Detects rubber-stamping (>95% approval) and extraction collapse (<20% approval) — writes PATTERN entries to learnings automatically.
Telegram Multi-Turn History New
20-message per-chatId history in Cache (24h TTL). Injected into AgentContext. /new clears it. Last 4 messages used as insight extraction context.
What Makes M.U.S.E. Unique
Adversarial Evaluator
A second LLM judges every insight batch before storage. Score < 0.4 = rejected. Prevents self-confirmation bias from corrupting memory.
Belief Map
Silent extraction of limiting and empowering beliefs from natural conversation. A private coaching layer that guides tone without the user ever seeing it.
Adaptive Fidelity
Scratch notes ranked by importance × recency × Jaccard relevance. The prompt always gets the most contextually relevant notes — not just the newest.
Skill Compounding
Repeated insight patterns (5× in 30 days) are synthesized into candidate AgentSkills. The system literally grows new capabilities from usage patterns.
Biological Memory
Mimics human memory consolidation — short-term notes are periodically condensed into long-term summaries by the Gardener.
Private Knowledge Guardian
The only agent with access to private RAG sources and infra runbooks. Acts as a filtered gateway to other agents.
Dynamic Prompt Budgets
Enforces % based character budgets: 40% Brand Bible · 15% Scratch · 30% RAG — preventing context overflow while maximizing info density.
Triple-Store Pipeline
Every insight is stored in the filesystem, vector DB, and optionally Google Drive — simultaneously. No insight is ever lost.
Full Brand Isolation
Memory files, vectors, cache keys, Drive connections, API keys, conversation history — everything is fully brand-scoped.
Multi-Agent Hub
Hub agent referenced by SEO, WordPress, WebBuilder agents — shares only the context each one needs, keeping secrets safe.
Data Flow Architecture
Automated Insight Categories
Every conversation is mined for these 6 categories. Confidence ≥ 0.6 required. Adversarial evaluator score ≥ 0.4 required.
Quality Tracking
Newquality_stats.json per brand tracks approval rates over time. Anomalies trigger automatic PATTERN entries in learnings.
9 Marketing Modes
M.U.S.E. auto-detects your intent from keywords and shifts its entire persona to match. Creative modes get full belief coaching. Tech modes get clean execution context.
Strategy
Market-fit, positioning, KPIs
Copy
Headlines, taglines, CTAs
Social
Short-form posts, hashtags
Campaigns, subject lines
SEO
Keywords, meta, on-page
PPC
Ad copy, budget, testing
Growth
Funnel optimization, experiments
Creative
Tone, visual direction, mood
General
Strategic, practical, actionable