Intelligent Search / Agentic AI · Knowledge · production
MythOS
AI-native knowledge platform with agent-addressable memory.
Living Markdown library where notes become a single source of truth—searchable, citable, and writable by AI agents over MCP and REST—so every model session starts with your memory, not a blank context window.
Problem
Knowledge is scattered across notebooks, chats, and duplicated CLAUDE.md files. Every n…
System
A RAG-centered knowledge graph of linked memos with semantic retrieval, memo-level cita…
Role
AI Architecture & RAG Systems
Status
production
Outcome
MCP Native agent protocol
My role
AI Architecture & RAG Systems
- AI Architecture
- Retrieval
- Agent Orchestration
- Integration
- Productionization
Period / 2024–Present
Agent-addressable single source of truth
INTENT
Plays once · tap to replay
Problem
Knowledge is scattered across notebooks, chats, and duplicated CLAUDE.md files. Every new AI session forces re-explanation, and context drifts across tools.
System
A RAG-centered knowledge graph of linked memos with semantic retrieval, memo-level citations, and an MCP/agent API so Claude, Cursor, ChatGPT, and automations read and write one library.
Architecture
Authoring and agent clients write into a linked memo graph; embeddings power semantic retrieval; MCP/REST expose search, CRUD, and grounded chat with citation-backed answers.
- Author / Editor → Memo Graph
- AI Clients → MCP / Agent API
- MCP / Agent API → Memo Graph
- Memo Graph → Embeddings
- Embeddings → Semantic Retrieval
- Semantic Retrieval → Cited Answer
- MCP / Agent API → Semantic Retrieval
- Memo Graph → Publish / Automations
Execution path plays once on view · hover a node or tap the canvas to replay
System anatomy
Architecture
Memo graph + retrieval layer + agent protocol surface.
- Living documents
- MCP server
- Agent REST API
- Discovery (/llms.txt)
Engineering decisions
Decision
Memory outside any single model
Constraint
Model vendors and chat UIs churn; personal knowledge must survive migrations.
Approach
A model-agnostic library with MCP and REST keeps retrieval and writes portable across Claude, ChatGPT, Cursor, and future clients.
Result
One source of truth addressable by any HTTP/MCP agent.
Decision
Living documents over static dumps
Constraint
Paste-into-chat context goes stale the moment knowledge changes.
Approach
Linked memos with @mentions and #tags form a graph agents can update once for every downstream session.
Result
Shared, updatable context without token-bloated duplication.
Decision
Citation-backed RAG answers
Constraint
Library chat must not invent knowledge detached from source memos.
Approach
Semantic retrieval returns answers in the author's words with memo-level citations for auditability.
Result
Grounded Q&A over personal and team libraries.
Decision
Permission-aware agent surface
Constraint
Private memos must stay private when agents are connected.
Approach
MCP/OAuth and API keys inherit per-memo permissions so agents never broaden access by default.
Result
Safe co-authoring with private-by-default memos.
Reliability
- Private-by-default memosACTIVE
- Permission tagsACTIVE
- Markdown export/portabilityACTIVE
- BYO API keysACTIVE
- Agent-scoped authACTIVE
Evaluation
Outcome / Results
- One living library addressable by every MCP-capable client without duplicating context files.
- Semantic search with memo-level citations grounded in the author's own writing.
- Agents can create, update, and retrieve memos mid-conversation with permission-aware access.
Stack
Product visuals
