Memoh is a self-hosted, multi-agent AI platform that orchestrates persistent bot agents across multiple communication channels (Telegram, Discord, Slack, WeChat, LINE, etc.) and runtime backends (native LLM, Claude Code, Codex, Grok, ACP). The system provides each agent with an isolated container workspace (containerd/Docker) including filesystem, desktop, browser, network overlay, and long-term memory (pgvector graph store). Agents run 24/7, can be scheduled, and support tool execution, MCP integration, and skill/app installation from a built-in marketplace. The architecture is a full-stack monorepo: a Go backend (Echo HTTP server, gRPC internal RPC, PostgreSQL + pgvector) with 160+ schema migrations, a Vue 3 web frontend (Pinia stores, i18n, memory graph visualization), an Electron desktop app, and TypeScript SDK/runtime packages. The backend handles session lifecycle, context compaction (LLM-based summarization with quality gates), tool approval flows, user input requests, and multi-tenant row-level security. The frontend provides chat UI, bot management, onboarding, keyboard shortcuts, and a supermarket for apps/skills. Key data flows: inbound messages from channel adapters are normalized and routed to the agent runtime, which builds a context fragment tree (collectors, selectors, placers, renderers), invokes the LLM provider with retry budgets, and streams responses back through SSE/WebSocket to the web UI or channel adapters. Memory extraction and compaction run asynchronously, updating the graph store and compacting history under token budgets. The system emphasizes durability with fencing tokens, decision transactions, and resume-after-interrupt capabilities.