RemindAI Model Agent Workflow

User request through protocol adaptation, AgentLoop cycle, and multi-component orchestration

Request Turn 1 — Tool Use Context Enrichment Final Response send input submit task chat completion tool_call detected execute read/exec/search tool output semantic search relevant memories re-prompt with tools+context final answer render markdown display result store conversation User Input channel UI Layer Flutter desktop AgentLoop Core orchestrator LLM API OpenAI / Anthropic / Gemini ToolShell Files · Exec · rg/fd/rtk Memory Qdrant · SQLite Legend request return security async trace

AgentLoop Core

  • • The AgentLoop is the central orchestrator — it mediates all component interactions
  • • Each loop iteration: prompt LLM → parse tool_calls → execute → collect results → re-prompt
  • • Activations make ownership duration visible across the timeline

Protocol Adaptation

  • • Three LLM APIs (OpenAI / Anthropic / Gemini) are abstracted behind a single interface
  • • Emphasis arrows mark the main request path; returns are quieter
  • • The same loop pattern applies regardless of which LLM backend is selected

Tool + Memory Augmentation

  • • ToolShell provides file sandbox, code exec, and CLI tools without extra installs
  • • Memory layer uses Qdrant semantic search + SQLite backup for persistent cross-session recall
  • • Both augment the LLM context before the next API call

Async Memory Store

  • • Conversation storage is non-blocking (dashed arrow)
  • • It happens after the user-facing response completes
  • • Separates observability from user latency

Sequence diagram • Built with Archify • Press T for theme and E for export