Votee AI and Beever AI Open-Source Beever Atlas for Chat-to-Wiki Conversion

Hong Kong-headquartered Votee AI and its Toronto-based research lab Beever AI have open-sourced Beever Atlas — an LLM Knowledge Base that transforms unstructured team chat into a structured, searchable wiki. The announcement, made May 8, 2026, positions the tool as a direct response to Andrej Karpathy's viral call for "an incredible new product" to solve LLM memory problems.

Beever Atlas automatically ingests conversations from Telegram, Discord, Mattermost, Microsoft Teams, and Slack, then distills them into a Neo4j knowledge graph with typed entity relationships between people, projects, technologies, and decisions. The system ships in two editions: an Apache 2.0 Open Source Edition for individuals, and an Enterprise Edition for regulated organizations like banks and government agencies.

According to the official press release, the platform addresses what co-founder and CEO Pak-Sun Ting calls "conversational knowledge loss" — the silent liability where organizational knowledge lives and dies in ephemeral chat streams.

The technical architecture makes a deliberate bet against vector similarity search. Jacky Chan, Co-Founder and CTO of Votee AI (who previously developed the first fully pre-trained open-source Cantonese LLM), stated: "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."

Beever Atlas runs as a Docker Compose stack with three services (backend, bot, frontend) backed by four data stores: Weaviate, Neo4j, MongoDB, and Redis. The dual-memory architecture combines a 3-tier semantic store for fast hybrid search with a graph store that extracts entities and relationships. This design means the expensive distillation work happens once at ingestion, while queries hit compact, pre-digested context.

The platform ships with a native MCP (Model Context Protocol) server, allowing AI assistants like Cursor, AWS Kiro, and Qwen Code to query team knowledge directly. OpenClaw and Hermes Agent integration is scheduled for Q2 2026, making it among the first MCP-native knowledge backends purpose-tuned for these workflows (a problem that has plagued developers for years, frankly).

Security and sovereignty are baked into the design. Beever Atlas runs entirely in customer environments with zero telemetry, AES-256-GCM encryption at rest, and private channels filtered by default. Teams bring their own LLM via LiteLLM — running locally through Ollama (Gemma, Qwen, Llama) or via 100+ supported cloud providers. This matters for organizations where knowledge is too sensitive for third-party cloud processing.

The Enterprise Edition extends the open-source core with five capabilities for regulated environments, including permission mirroring that prevents AI tools from accidentally leaking private HR or salary information to junior employees. The open-source edition targets solo developers, content creators, and researchers running personal knowledge management against their own chat workspaces.

Installation requires two free API keys: a Google API key for Gemini (extraction, entity graph, answers) and a Jina API key for embeddings. The GitHub repository includes a 30-second seeded demo with pre-computed fixtures — no API keys required for initial testing.

The wiki generation pipeline runs in three phases: Gather (pulls semantic clusters from Weaviate and entities from Neo4j), Compile (per-page LLM calls in parallel with deterministic prompts), and Cache (compiled pages land in MongoDB with dirty flags for incremental syncs). Quality control rules run daily — clusters with coherence below 0.4 re-cluster from scratch, while those with 100+ members split into sub-clusters.

Unlike Karpathy's prototype, which starts with curated file ingestion and relies on Obsidian and an LLM coding agent, Beever Atlas takes team chat as its starting point. The system handles multimodal intelligence — text, images, voice, video, and PDFs unified in one searchable memory layer — and supports multi-user architecture rather than single-user only.

The physical experience differs from typical RAG systems. Instead of scrolling through three months of chat history, new teammates can browse a distilled wiki with topic pages, entity graphs, decisions, and citations. Every wiki claim links back to source messages, making answers auditable all the way down to the original Slack or Discord thread.

Votee AI operates across Asia with headquarters in Hong Kong and offices in Toronto, Ho Chi Minh City, and Kuala Lumpur. The company previously validated its platform in the Hong Kong Monetary Authority's highly regulated fintech sandbox, which informs the Enterprise Edition's security architecture.

Whether organizations actually deploy this at scale remains the real question. The tool solves a genuine problem — knowledge loss in chat — but requires teams to invest in self-hosting infrastructure and API key management. For regulated industries, the sovereignty angle is compelling. For everyone else, the friction of setup might outweigh the benefits of having a wiki that writes itself.

Time will tell if Beever Atlas becomes the memory layer every downstream AI agent needs, or another well-intentioned open-source project that gathers dust in a GitHub repo. The code is public, the architecture is sound, and the problem is real. Whether users actually pay for the Enterprise Edition is the only metric that matters.

Artūras Malašauskas is an AI Systems Integrator with 20+ years of production-grade web engineering experience. He has designed, shipped, and scaled enterprise Python/PHP systems for logistics, SaaS, and public-sector clients. For the past year, he has focused exclusively on AI integrations: deploying open-source LLMs, building generative media pipelines (image, audio, video), and engineering multi-agent workflows for real production environments. His standard: reproducibility, security, cost-efficient inference—no vaporware. He documents and evaluates emerging AI tooling, separating verified capabilities from marketing noise. Technical editor at: muza-ai.eu, ai-verslas.lt, ai-naujinos.lt Connect on LinkedIn