Beever Atlas

Beever Atlas is one of the few genuinely interesting entries in the "team LLM wiki" category, and it earns that mostly by being honest about what it is: a self-hosted pipeline, not a hosted chatbot with a memory tab. The architecture is the product. Chat messages flow through a 6-stage pipeline (sync, extract, validate, store, cluster, wiki). The validate stage is the design decision we like most — extractions below a 0.5 confidence threshold are dropped rather than written, which means your wiki is smaller but less polluted by LLM hallucination than a naive "summarize every channel" approach. The dual-memory split is the second decision we like: Weaviate handles the ~80% of queries where you just want semantic similarity in under 200ms, and Neo4j handles the ~20% where you actually need relationships — "who decided X," "how did our approach to auth evolve," "which decisions touched the API design." Graph traversal is what makes temporal and multi-hop questions answerable at all, and almost no competitor in this space does it for chat history. The MCP surface is a smart bet. Exposing beever_ask, beever_search, beever_graph, and beever_wiki means Atlas drops into Claude Desktop or Claude Code as a tool rather than forcing you into a new UI, and the REST + Docker combo covers the rest. Auto-generated wiki pages with citations back to the source Slack or Discord message are the feature that will sell this internally — a decision doc that cites the thread where the decision happened is materially more trustworthy than a Notion page someone forgot to update. Where it gets harder: v0.1.0 is early, and the docs are candid that features and stability will evolve. Self-hosting means operating Weaviate and Neo4j, not just one container — that's real work, and the docs we scraped don't publish rate limits or concurrency-scaling figures, so you're sizing the stack yourself for large workspaces. The open-source release also doesn't include the governance layer (SSO/SAML, audit logs, RBAC, spend controls) that the Enterprise tier lists. Practical fit: engineering teams that already run Docker and want architectural-decision recall; product teams clustering Discord and Slack feedback into themes with source links; remote teams reconstructing what happened while they were asleep; open-source communities summarizing RFC and plugin consensus. Poor fit: anyone who wants a managed SaaS with no servers, anyone whose chat lives in email or a tool outside the four supported platforms, and anyone buying this as a personal second brain — it's team-scoped. Compare against Mem0 (managed memory layer, less wiki output), Khoj (more general personal/team assistant), Obsidian+LLM (notes-first, no chat ingestion), and Notion AI (managed, no graph, no self-host). Atlas's moat is the combination — open source, self-hosted, chat ingestion, dual memory, MCP — and it's the only one on that list with all five.