🧠 AgentOS Cortex — Knowledge Compiler

Your AI's Long-Term Memory

Cortex turns disposable AI research into a persistent, compounding knowledge base. No vector databases. No RAG pipelines. Just structured Markdown compiled and maintained by your LLM.

Cortex = An LLM-compiled wiki layer that sits between your raw research sources and your agents. Every document, insight, and decision is transformed into human-readable, interlinked Markdown — making your AI's knowledge base inspectable, editable, and self-healing.

100

– 10K docs sweet spot

Embeddings needed

Self-
healing

LLM linting

Why RAG Falls Short at Personal Scale

the problem

❌ RAG — Vector Retrieval

1. Chunk documents into fixed-size fragments

2. Embed each chunk into high-dimensional vectors

3. Store in a vector database (Qdrant, Pinecone, Weaviate)

4. At query time: embed the question, find nearest-neighbor chunks

5. Hope the retrieved chunks are actually relevant

✗ Opaque embeddings — you can't inspect what was retrieved or why

✗ Retrieval noise — nearest neighbor ≠ most relevant

✗ Static index — adding new docs requires re-embedding

✗ Chunking destroys context — a paragraph without its surrounding doc loses meaning

✗ Black box — no way to audit, edit, or fix what the system "knows"

✅ Cortex — LLM Wiki

1. Dump raw materials into raw/ directory

2. LLM reads raw content, writes structured encyclopedia articles

3. Articles live as plain .md files with backlinks and index files

4. At query time: LLM reads relevant articles directly — no vector lookup

5. Results feed back into the wiki, compounding over time

✓ Fully inspectable — open any .md file and read it

✓ Explicit backlinks — every connection is deliberate

✓ Editable — fix or extend any article directly

✓ Self-healing — LLM linting finds and fixes inconsistencies

✓ Zero infrastructure — local files + any LLM

The Scale Insight: RAG is designed for millions of documents where humans can't read everything. At 100–10,000 high-signal documents — personal research, business intelligence, market analysis — an LLM can read and synthesize everything. Skip the embedding pipeline. Compile a wiki instead.

The Cortex Architecture

compile pipeline

🌐 Sources Perplexity · Vane · Obsidian Web Clipper · Business Docs · Research Papers

↓

📁 raw/ directory Unprocessed source material — .md, .pdf, .txt, web clips

↓

🧠 LLM Compiler Reads raw/ → writes structured wiki articles, backlinks, index files

↓

🗂️ cortex/wiki/ directory .md files · backlinks · category index · master index.md

↙ ↓ ↘

🔍 Query Interface LLM reads wiki directly — no vector search

📊 Obsidian Human viewing · graph visualization · backlink explorer

♻️ Lint & Heal Periodic LLM health checks · gap-filling · consistency

RAG vs Cortex — Full Comparison

7 dimensions

Dimension	RAG (Vector Embeddings)	Cortex (LLM Wiki)
Storage	Opaque vector embeddings — unreadable binary blobs	Human-readable .md files — open in any text editor
Retrieval	Semantic similarity via nearest-neighbor search	Explicit backlinks + index files — no guessing
Traceability	Low — black box, can't audit what was retrieved	High — every claim links to a specific source file
Maintenance	Static index requires full re-embedding to update	Self-healing via periodic LLM linting passes
Scale Sweet Spot	Millions of documents (enterprise, web-scale)	100–10,000 high-signal documents (personal/team)
Editability	Cannot edit embeddings — must re-index from scratch	Edit any .md file directly — changes instant
Infrastructure	Vector DB (Qdrant/Pinecone) + embedding pipeline + hosting	Local .md files + any LLM — zero servers needed

The Compile Pipeline — Step by Step

5 stages

Data Ingest — Filling raw/

Raw materials from all intelligence sources land in the raw/ directory. No processing yet — just collection.

Perplexity research outputs exported as Markdown
Vane private company analysis and intelligence reports
Web articles captured via Obsidian Web Clipper (one-click .md conversion)
Business documents, earnings calls, regulatory filings
Research papers and whitepapers (.pdf converted to .md)
Meeting notes, decision logs, personal observations

Perplexity Vane Obsidian Clipper PDF → md

LLM Compilation — The Core Innovation

The LLM reads all raw material and writes a structured, encyclopedia-style wiki. This is not summarization — it is knowledge synthesis and organization.

Generates article-length summaries of all source material
Identifies key concepts and writes dedicated encyclopedia-style articles for each
Creates explicit [[backlinks]] between related articles (Obsidian-compatible syntax)
Builds and maintains category index files (e.g. markets/index.md)
Builds and maintains a master index.md linking all articles
Categorizes content into the directory structure automatically
Assigns confidence scores and flags unverified claims

Core Innovation Claude / GPT-4o wiki compilation

Active Linting — Self-Healing Health Checks

On a scheduled cadence (daily or weekly), the LLM runs health-check passes over the entire wiki to maintain integrity and discover new connections.

Detects contradictions between articles (e.g. two articles claiming different market share figures)
Imputes missing data by triggering web searches via Perplexity or Vane
Discovers new connections between previously unlinked concepts — adds backlinks
Suggests new articles to write based on gaps in coverage
Updates stale articles when newer source material is added to raw/
Incrementally improves confidence scores and data integrity over time

Scheduled Perplexity Incremental

Query — Ask the Wiki Directly

Complex questions answered by pointing the LLM at relevant wiki articles. No embedding lookup — the LLM reads the actual articles, with full context and backlinks preserved.

LLM reads the master index.md to find relevant articles
Loads the specific .md articles into context window
Answers the question with explicit article citations
Can traverse backlinks to pull in related context automatically
At 100–10K articles, the entire wiki fits within or near modern context windows

No vector DB Full context Cited answers

Output Loop — Every Exploration Compounds

Results from queries, agent tasks, and new research are filed back into the wiki. Each run makes the next run more informed.

Query answers filed as new articles or appended to existing ones
Agent task outputs (Manus executions) compiled into decisions/ and projects/
New raw/ files trigger incremental recompilation of affected articles
Over time: a continuously improving, compounding knowledge base

Compounding Feedback loop Manus outputs

Cortex in AgentOS — Integration Points

cross-layer connections

L4 Intelligence → Cortex

Perplexity Cloud Research

→ feeds raw/

Every Perplexity deep research session exports to raw/. The LLM compiler ingests it on the next compile pass, adding the research to the wiki as structured articles rather than one-time outputs.

L4 Intelligence → Cortex

Vane Private Analysis

→ feeds raw/

Vane's private company and market intelligence — unavailable to public search — lands in raw/ and gets compiled into the competitors/ and markets/ sections of the wiki.

L6 Memory → Cortex boundary

Qdrant vs Cortex Division

↔ complementary

Qdrant handles ephemeral session memory — what happened in this conversation, this task, this hour. Cortex handles persistent knowledge — what we know about markets, competitors, products, decisions across all time.

L3 Planner → Cortex

Manus Pre-Task Lookup

→ reads wiki before acting

Before Manus starts any significant task, it queries Cortex first: "What do we already know about this?" Prevents redundant research and surfaces relevant prior decisions, competitors, and context automatically.

L5 Execution → Cortex

Task Output Filing

→ writes to wiki

Completed Manus task outputs — analysis, decisions, code reviews, business assessments — are compiled back into Cortex. Every execution adds to institutional memory instead of disappearing into chat history.

L7 Culture → Cortex

Prompt Library & Decisions

→ stored in prompts/ + decisions/

Reusable prompts, standard operating procedures, and business decisions with full rationale are stored in Cortex. The AI Shop's accumulated judgment is persisted, not just in people's heads.

Recommended Vault Structure

AI Shop configuration

The Obsidian vault doubles as both your Cortex directory and human-readable knowledge base. The structure below is optimized for a product business doing active market research and competitive intelligence.

cortex/ ├── raw/ ← Perplexity exports, Vane reports, web clips — unprocessed ├── wiki/ ← LLM-compiled articles (the knowledge base) │ ├── markets/ ← Market sizing, trends, sector analysis articles │ ├── competitors/← One article per competitor, auto-updated on new raw/ │ ├── financials/← Earnings summaries, valuation models, sector benchmarks │ ├── products/ ← Product knowledge, feature comparisons, positioning │ └── strategy/ ← Strategic decisions with full context and rationale ├── projects/ ← Per-product knowledge and project history ├── prompts/ ← Reusable prompt library, SOP templates ├── decisions/ ← Business decisions with rationale, date, and context └── index.md ← Master index, auto-maintained by LLM compiler

COMPILE TARGET

wiki/ is the output of the compiler. Never manually edit wiki/ articles — let the LLM own them. Edit raw/ instead.

HUMAN EDITABLE

decisions/ and prompts/ are human-authored. The LLM indexes and links them, but doesn't overwrite them.

OBSIDIAN LAYER

Open the entire cortex/ directory as an Obsidian vault. Graph view shows the knowledge network visually.

🔬 The Karpathy Insight — April 2026

"A large fraction of my recent token throughput goes less into manipulating code, and more into manipulating knowledge."

— Andrej Karpathy, "LLM Knowledge Bases" (April 2026)

Karpathy's insight is that the LLM's role is shifting from code assistant to knowledge compiler. His personal implementation: roughly 100 articles, ~400K words of accumulated research — all LLM-compiled, maintained, and queryable. Not a vector database. A wiki.

LLMs as Compilers

Use the LLM's reasoning ability to organize knowledge, not just retrieve it. Compilation produces something better than its inputs.

Knowledge over Code

As AI handles more code, the limiting resource becomes organized, high-quality knowledge that agents can reason over.

~100 Articles Scale

Karpathy's personal wiki at ~100 articles / ~400K words is fully manageable by a single LLM call. No infrastructure needed.

Compounding Returns

Unlike RAG indexes that are static, a compiled wiki improves with each linting pass. The knowledge base compounds with use.

Implementation Roadmap

6 build steps

Step 01

Initialize Obsidian Vault

Create the cortex/ directory structure above. Open it as an Obsidian vault. Enable the Graph View and Backlinks plugins. This is both your Cortex working directory and your human-readable knowledge interface.

Obsidian Setup

Step 02

Configure Obsidian Web Clipper

Install the Obsidian Web Clipper browser extension. Configure it to save directly to cortex/raw/ with auto-generated frontmatter (source URL, date, tags). One-click capture of any web content to Markdown.

Obsidian Web Clipper Browser

Step 03

Build compile.py CLI

Python script that reads all files in raw/, sends them to the LLM with a compilation prompt, and writes the resulting wiki articles to wiki/. Include incremental mode (only recompile articles affected by new raw/ files) and a force-full flag.

Python Core Tool LLM API

Step 04

Build lint.py CLI

Scheduled health check script. Reads all wiki/ articles, sends to LLM with a linting prompt, receives a structured report of: contradictions found, gaps identified, new backlinks suggested, articles to write. Optionally auto-applies non-destructive fixes.

Scheduled Python Self-healing

Step 05

Build query.py CLI

Interactive query interface. Reads index.md to identify relevant articles for the question, loads those articles into context, queries the LLM with the loaded context + question, returns cited answer. Optionally files the answer back into wiki/.

Python Interactive Cited

Step 06

Integrate with Manus Orchestrator

Register compile, lint, and query as available tools in the Manus orchestrator. Manus can then: pre-task lookup ("what do we know about X?"), post-task filing ("file this result to wiki/"), and trigger lint passes when the wiki grows stale.

Manus AgentOS Automated

→