← Redis
Cole Medin
I Love the Karpathy LLM Wiki but it Doesn't Scale. Here's What Does.
⏱ 19 min video · 3 min read9 Jul 2026Worth watching
TL;DR
Cole Medin argues that markdown-based personal agents (like Karpathy LLM Wikis) don't scale to production, and walks through the architecture needed for production agents: a database-backed context retrieval layer and persistent user memory. He uses Redis Iris as a live demonstration platform.
Key points
1
Personal agents using markdown files (LLM Wikis, Obsidian, Hermes) are great for individuals but fundamentally cannot scale to production due to cost, governance, and retrieval limitations.
2
Production agents require two core components: a context retriever (giving the agent structured access to business data) and agent memory (short-term and long-term memory per user).
3
Redis Iris provides both via an auto-generated MCP server for context retrieval and an API-based agent memory service that automatically promotes key facts from session memory to long-term vector storage.
4
The context retriever auto-generates MCP tools based on defined entities and schemas, allowing the agent to filter and search efficiently without needing to know the database schema itself.
5
Agent memory uses vector/semantic search so individual users can accumulate millions of memories and the agent still retrieves the most relevant ones per conversation.
Actionable insights
→
Use markdown-based LLM Wikis only for personal agents; the moment you ship to other users, switch to a database-backed architecture with structured context retrieval.
→
Define your data entities and schema in a context retrieval service (like Redis Iris) to auto-generate MCP tools — this removes the need for the agent to infer schema and dramatically cuts token usage.
→
Implement both short-term (session) and long-term (promoted vector) memory per user so your agent builds intelligence over time without re-reading entire conversation histories.
→
Use Pydantic AI over coding-agent SDKs (Claude Code SDK, Codex SDK) for production agents — it is faster and more token-efficient for non-coding agentic tasks.
→
Attach a context retriever MCP server to your agent by passing the documentation to Claude Code and letting it one-shot the integration, saving setup time.
Notable quotes
“Almost anything that's really providing real business value is an agent that is shipped as a part of a platform to a production environment with other people logging in and talking to the agent.”
“Coding agent SDKs, like the Claude agent SDK or Codex SDK, I know they're very popular now, but they're slow because they're made for longer agentic coding tasks, and they're also more token heavy.”
“I don't even think it spent a thousand tokens to get this information.”
Worth watching?
✅
Worth watching the full video?
Watch if you are building or planning to ship AI agents to real users — the live Redis Iris demo is clarifying, but the architectural principles apply universally and could save you from a costly rebuild later.
Saved you some time? The creator still deserves a like.
Found this by searching? That's the hard way.
summree watches the AI channels you follow and sends you the summary the moment a video drops. Nothing gets past you. From £4/month.
Try it free