Git-native memory, a typed knowledge graph, and MCP-powered retrieval on top of plain Markdown — so agents build, evolve, and reuse persistent knowledge instead of re-deriving it on every query.
Knowledge that compounds.
Traditional RAG retrieves documents and forgets. Cortex builds a persistent, evolving knowledge layer that grows over time — and becomes more valuable the more agents and humans interact with it.
RAG chunks and re-indexes the same sources on every query. Nothing is learned, nothing is kept, nothing connects.
Insights vanish when the context window closes. Agents re-derive what they already worked out yesterday — and can't build on each other.
Every read, write, ingest and compile is logged and graph-linked. Files survive sessions, deployments and model upgrades — diff-friendly Markdown, no database.
Backed by a typed knowledge graph, hybrid search, and a MarkItDown-powered file compiler. Drop files in. Let agents read, write, search, link and compile knowledge.
Every read, write, ingest and compile event is logged and graph-linked — the vault learns as it's used.
Wikilinks, tags and typed edges, maintained automatically. vault_context builds a ranked subgraph from any seed query.
Markdown notes with typed nodes — note, agent_def, session, daily, feedback.
Hybrid search — BM25 + vector + re-ranking — with keyword and semantic modes, cached with a configurable TTL.
One vault, every MCP-compatible client — streamable HTTP or stdio, mirrored 1:1 by a REST API at /api/v1/.
The vault is a plain directory of Markdown — diff-friendly out of the box. Zero database; everything persists as Markdown + JSON on your filesystem.
The same vault is reachable over MCP and REST, kept in sync by a real-time filesystem watcher. Everything below the surface is plain Markdown plus a typed graph.
PDF, DOCX, PPTX, XLSX, HTML, images and more are converted to Markdown via MarkItDown. Bulk-ingest hundreds at once with SHA-256 dedup.
Raw sources compile into interlinked wiki articles. An LLM is used only to detect cross-references between articles — file conversion needs no API key.
A NetworkX engine maintains wikilinks, tags and typed edges on every file change. The vault watcher keeps the graph in sync automatically.
Agents query over MCP; vault_context returns a ranked subgraph. Every operation mirrors into a greppable daily activity log.
Every tool is exposed over MCP and mirrored 1:1 in the REST API. Plug into any AI client that speaks the Model Context Protocol.
Read a note by path.
Create or update a note.
Search the vault — keyword, semantic or hybrid.
Create, query or delete graph edges.
Search → graph traversal → ranked subgraph.
Ingest raw content or binary files.
Compile raw sources into wiki Markdown.
Submit feedback on vault quality.
List feedback note metadata.
RAG retrieves chunks. GraphRAG adds structure but no memory. Cortex is a persistent, Git-versioned, MCP-native knowledge layer.
| Capability | PULSE8.ai Cortex | Traditional RAG | GraphRAG |
|---|---|---|---|
| Persistent knowledge | ✓ | ✕ | ~ |
| Markdown-native storage | ✓ | ✕ | ✕ |
| MCP-compatible out of the box | ✓ | ✕ | ✕ |
| Knowledge graph | ✓ | ✕ | ✓ |
| Git versioning | ✓ | ✕ | ✕ |
| Agent memory layer | ✓ | ✕ | ~ |
| Human + AI collaboration | ✓ | ✕ | ~ |
| Continuous knowledge evolution | ✓ | ✕ | ~ |
| Zero database required | ✓ | ✕ | ✕ |
Start free and self-hosted. Bring in managed deployment, single sign-on, compliance guardrails and support when Cortex moves into production.
The full knowledge layer, on your own infrastructure. No database, no lock-in.
Free · Apache 2.0
For regulated teams running Cortex in production — part of the PULSE8.AI platform.
Custom · let's talk
Everything in Open Source, plus:
The vault is just files, so humans and agents collaborate on the same knowledge. No proprietary format, no lock-in.
Cortex requires Docker. An OpenRouter key is optional — needed only for LLM cross-referencing between articles. File conversion works with no key at all.
Grab the source from GitHub and step in.
start.sh builds Cortex (API + MCP on :8420) and QMD (search on :3100), waits for health checks, and you're ready.
Point Claude Desktop — or any MCP client — at http://localhost:8420/mcp/.
# 1 · clone $ git clone https://github.com/synpulse8-opensource/\ pulse8-ai-cortex-knowledge-vault.git $ cd cortex-knowledge-vault # 2 · launch (Cortex :8420 + QMD :3100) $ ./scripts/start.sh ✓ cortex healthy → http://localhost:8420 ✓ qmd healthy → http://localhost:3100 # 3 · connect your MCP client → http://localhost:8420/mcp/
A persistent, compounding knowledge base maintained by LLMs — compiled and interlinked incrementally rather than re-derived from raw documents on every query. The direct inspiration for Cortex's architecture.
The on-device engine powering all full-text and hybrid search in Cortex — combining BM25, vector search and LLM re-ranking, all running locally.
The file-to-Markdown converter behind the Cortex compiler — turning PDF, Office docs, HTML and images into structured Markdown for ingestion.
Star the repo, open a discussion, and help build the knowledge layer for agentic AI.