Andrej Karpathy, OpenAI's founding member and former Tesla AI director, posted a viral call for a new type of AI product, and two companies have now delivered. Hong Kong-based Votee AI and its Toronto research lab Beever AI have open-sourced Beever Atlas, an LLM (large language model) knowledge base that automatically transforms team conversations from Slack, Microsoft Teams, Discord, Mattermost, and Telegram into structured, searchable organizational memory.

What Problem Is Beever Atlas Actually Solving?

Karpathy's original observation was straightforward but powerful: large language models need structured, evolving knowledge to work effectively, not just raw context windows or basic similarity search. He concluded his viral post with a direct challenge to the industry: "I think there is room here for an incredible new product instead of a hacky collection of scripts".

The core issue is that most organizational knowledge lives and dies in unstructured conversations. Teams spend hours discussing decisions, sharing insights, and solving problems in chat platforms, but that knowledge becomes nearly impossible to retrieve or leverage later. Beever Atlas treats this perishable resource as a compounding organizational asset by automatically converting chat into a Neo4j knowledge graph, auto-generated wiki, and memory layer that AI assistants can query directly.

"Every growing organization faces the same silent liability: conversational knowledge loss. Beever Atlas turns this perishable resource into a compounding organizational asset," said Pak-Sun Ting, Co-Founder and CEO of Votee AI.

Pak-Sun Ting, Co-Founder and CEO of Votee AI

How Does Beever Atlas Differ From Karpathy's Original Vision?

Karpathy's prototype started with manual file ingestion through Obsidian and relied on Claude Code or Codex to organize knowledge. It was single-user and largely manual. Beever Atlas takes a fundamentally different approach by starting with team chat as the primary data source, making it inherently multi-user and automated.

The two products diverge in several critical ways:

• Data Source: Beever Atlas ingests directly from Telegram, Discord, Mattermost, Microsoft Teams, and Slack conversations, rather than requiring manual file uploads.
• User Interface: Beever Atlas offers a zero-install web UI, eliminating the need for Obsidian or command-line tools that Karpathy's prototype required.
• Knowledge Structure: Beever Atlas builds a full Neo4j knowledge graph with typed entity relationships between people, projects, technologies, and decisions, not just text-only cross-references.
• Multimodal Support: The product handles text, images, voice, video, and PDFs in a unified searchable memory layer, whereas Karpathy's prototype was text-only.
• AI Agent Integration: Beever Atlas ships with native Model Context Protocol (MCP) server integration, allowing tools like Cursor, AWS Kiro, and Qwen Code to query team knowledge directly. Karpathy's prototype has no agent integration.

How to Deploy Beever Atlas for Your Organization

The product ships in two editions designed for different use cases and security requirements:

• Open Source Edition: Available under Apache 2.0 license for individuals, solo developers, content creators, and researchers managing personal knowledge against their own Telegram, Discord, or personal Slack and Teams workspaces. This version is free and self-hostable.
• Enterprise Edition: Purpose-built for banks, government agencies, and large organizations with high-security requirements. It extends the open-source core with permission mirroring, identity and multi-tenancy management, audit and compliance logging, and prompt-injection defense.
• Deployment Model: Beever Atlas runs entirely in customer environments as a Docker stack with zero telemetry. Teams bring their own LLM via LiteLLM, running locally through Ollama or via 100+ supported cloud providers.

The enterprise edition includes several security-focused features. Permission mirroring ensures that if a user lacks access to a private channel in Slack or Microsoft Teams, the AI cannot use information from that channel to answer their questions. Permission changes propagate in under 60 seconds, so when a user is removed from a project channel, the AI stops answering their questions about that project almost instantly.

"The key technical decision was to treat agent memory as a knowledge engineering problem, not a retrieval problem. Structure beats similarity, a typed graph of who works on what is more useful to an AI than vector search over a Slack archive," explained Jacky Chan, Co-Founder and CTO of Votee AI.

Jacky Chan, Co-Founder and CTO of Votee AI

Why Does Karpathy's Endorsement Matter for AI Infrastructure?

Karpathy's viral post reached tens of millions of impressions and crystallized a growing frustration in the AI community: current approaches to knowledge management for language models are fragmented and inefficient. His call for "an incredible new product" essentially set a benchmark for what the industry should build.

Beever Atlas's response demonstrates how Karpathy's observation identified a genuine market gap. The product addresses his core critique by building knowledge management from the ground up for team environments, where the bulk of organizational intelligence actually resides. The fact that two companies moved quickly to deliver a solution suggests this is a problem many organizations are actively trying to solve.

The product also signals a broader shift in how AI infrastructure companies are thinking about knowledge. Rather than treating knowledge as static documents to be ingested, Beever Atlas treats it as a living, evolving asset embedded in team conversations. This approach aligns with how modern organizations actually work and how AI assistants need to operate to be genuinely useful.

Upcoming integrations with OpenClaw and Hermes Agent, shipping in Q2 2026, will further expand the product's utility by allowing these AI tools to read and write to a user's Beever Atlas memory layer natively, making it among the first MCP-native knowledge backends purpose-tuned for these workflows.