1.5 KiB
2026-02-27 Session Notes
Signet Architecture Overview
Nicholai asked how Signet connects Claude Code, OpenCode, and OpenClaw. The conversation focused on Signet's role as a portable identity and memory layer that works across multiple AI coding platforms.
Key architectural insight: Signet doesn't replace any of these tools—it provides the stateful layer they all lack. Each tool is stateless and forgets between sessions; Signet captures context and reinjects it regardless of which frontend is used.
Core Connection Mechanisms
Single Source of Truth: All configuration lives at ~/.agents/. The AGENTS.md file is canonical; Signet auto-generates CLAUDE.md for Claude Code (with header "AUTO-GENERATED from ~/.agents/AGENTS.md"). OpenCode and OpenClaw read from the same directory.
Memory Continuity: The Signet daemon runs independently and provides memory through MCP tools (memory_search, memory_store, etc.). Since MCP is protocol-agnostic, any tool supporting MCP servers gets access to the same memory database. Sessions are summarized post-completion, facts are extracted and scored by importance, then embedded for hybrid search.
Secrets Management: Secrets stored once in Signet's vault, injected as env vars via secret_list and secret_exec MCP tools at runtime with output redacted. Works across all three platforms.
Tool-Agnostic Design: The user began asking about how Signet works regardless of MCP or skills support (the hooks-based fallback approach), but the conversation was cut short before that explanation was provided.