OpenPoke: Recreating Poke's Architecture (www.shloked.com)

After Poke — an iMessage AI assistant from the Interaction Company — went viral for its personality and utility, a leaked system prompt let a developer reverse-engineer its design and release OpenPoke: a stripped-down, local prototype that recreates Poke’s core architecture. The timing matters: Poke is one of the first post‑ChatGPT consumer AIs people actually rely on for daily work (email triage, reminders, lightweight automation), and OpenPoke surfaces the concrete engineering patterns that make that possible while also highlighting privacy and reproducibility trade-offs exposed by the prompt leak. Technically, OpenPoke swaps the single-threaded tool-call model for an orchestrated multi-agent system. A single Interaction Agent manages context, personality, and delegation; it spawns persistent Execution Agents on demand that run in parallel, own specific tasks (e.g., “Email to Alice”), keep full tool-call logs, and maintain long-lived memory. Agents chain atomic tools (Gmail ops, trigger management) into higher-level tasks (search, Notion/Linear MCP in Poke) and attach scheduled triggers backed by an SQL scheduler. Email monitoring runs continuously via LLM classifiers to surface important messages. Memory is multi-tiered: full recent context, compressed long-term summaries, and per-agent durable logs that preserve conversational continuity. The result is scalable parallelism, natural background automation, and richer continuity—patterns builders can steal, but with caveats around integration complexity and data protection.