The architecture behind your second brain
Kybernesis extracts facts, detects contradictions, builds entity profiles, and retrieves with precision—all while you sleep.
The Big Picture
Imagine having a brilliant assistant who reads everything you give them, extracts the key facts, builds profiles of every person and project, detects when information contradicts what it already knows, and can instantly recall exactly what you need—with precision, not noise.
That's Kybernesis. Not a document store—a knowledge engine.
From upload to intelligence
Here's how Kybernesis transforms your raw documents into structured, searchable knowledge.
Bring in your knowledge
Drop PDFs, docs, spreadsheets, or paste text directly. Connect Google Drive or Notion for automatic syncing. We handle OCR for scanned documents.
Smart extraction
Content is chunked intelligently (~500 words each), keeping ideas together at natural boundaries. Each chunk is embedded into 256-dimensional vectors to capture semantic meaning.
Four-layer storage
Memories live across vector store (256d semantic embeddings), structured database (metadata & relationships), fact store (atomic entity-attribute-value triples), and entity profiles (auto-built from aggregated facts).
The 8-step intelligence pipeline
Every hour, while you're working (or actually sleeping), our AI agent runs a pipeline that transforms raw documents into structured knowledge.
AI extracts semantic tags—entities, topics, and themes—not just keywords. A document about revenue projections gets tagged finance, Q4, revenue, strategic-planning.
AI reads each memory and extracts atomic facts: "Sarah's role is engineering lead", "Project Alpha launches April 15". Each fact is stored with entity, attribute, value, and confidence.
When new facts conflict with existing ones, the system detects contradictions and automatically supersedes outdated information. Your knowledge stays current without losing history.
Aggregates facts into living profiles for people, projects, organizations, and topics. Ask "What do I know about Sarah?" and the system has a ready profile with static facts, dynamic context, and related entities.
Memories get connected with typed edges: related, updates, extends, contradicts. Your project proposal automatically links to research notes, meeting summaries, and budget documents.
Recent, frequently-accessed, highly-connected memories stay in the hot tier. Older, isolated memories move to warm then archive. Nothing is deleted—just organized by relevance.
Beyond document search
Most memory systems store documents and search them. Kybernesis understands what your documents mean and tracks how knowledge evolves over time.
Relative dates like "next Thursday" or "last month" are resolved to absolute dates. Facts referencing past events are automatically marked as expired, so retrieval prioritizes current information. "The meeting tomorrow" from three weeks ago won't pollute today's results.
Every fact version is preserved with timestamps and supersession chains. When information changes, the old version is deprioritized—not deleted. You can always trace how knowledge evolved: who held what role, when plans changed, and what the latest truth is.
3-way hybrid retrieval
Three retrieval strategies work together to find the most relevant results with minimal noise.
Semantic similarity using 256d embeddings. Finds documents about the same concepts, even with different words.
Tag overlap, recency, access patterns, and tier priority. Structure meets meaning.
Direct text matching for names, dates, and specific terms that embeddings can miss.
Chunk Limiting
Returns only the top 2-3 most relevant chunks per memory, not entire documents. Less noise, more signal.
Relevance Floor
Results below a minimum relevance threshold are filtered out. No more padding results with irrelevant matches.
Fact-Aware Scoring
Memories with current (non-superseded) facts matching your query entities get boosted. Outdated facts get deprioritized.
Profile Injection
When your query mentions a known entity, the system includes their profile alongside search results—giving agents instant structured context.
Profile-powered AI agents
Create AI agents that use your memories, facts, and entity profiles as their knowledge base. They don't just search—they know.
Agents access auto-built profiles for instant context. Mention a person, project, or topic and the agent already knows the key facts.
Agents evolve through conversation. Say 'be more casual' and their persona updates instantly. Context persists across sessions.
Agents query your workspace memories via 3-way hybrid search with fact-aware scoring. Instant access to all your knowledge.
See your knowledge as a living graph
The Topology visualizes memories as nodes and relationships as edges. Watch clusters form around themes, discover unexpected connections, and see how entity profiles link your knowledge together.
Privacy & Security
Your memories are yours.
What makes this different
It extracts knowledge, not just documents
Most memory systems store and search raw text. Kybernesis extracts atomic facts, builds entity profiles, and creates a structured knowledge graph from your unstructured data.
It knows what's current
Contradiction detection and temporal grounding ensure you always get the latest information. When facts change, old versions are superseded—not deleted, just deprioritized.
It retrieves with precision
3-way hybrid scoring, chunk limiting, relevance floors, and fact-aware ranking mean you get the most relevant excerpts—not 40 chunks of loosely related text.
It gives agents real context
Entity profiles and structured facts mean your AI agents don't just search—they know. Ask about a person and the agent has their role, projects, and relationships at hand.
Your knowledge engine is ready
Upload a few documents. Watch the Sleep Agent extract facts, build profiles, and organize everything. Search using natural language. Create an agent that actually knows your context.