sacred-codex-pack

Imported from yogimathius/sacred-memory (sacred-codex-pack/AGENTS.md). Install upstream with npx skills add yogimathius/sacred-memory --skill sacred-codex-pack. Copyright stays with the author.

Agent Operational Instructions

This file defines cross‑agent rules for interacting with the sacred codex architecture. It should be imported or referenced by tool‑specific instruction files.

Core principles

• Treat canon content as immutable. Never modify or write to the canon directly.
• Use the memory harness (or MCP service) to retrieve context at session start. Do not search the file system directly.
• Save new information using /remember or /propose-canon commands so that it is captured by the harness and placed into working memory or the reflection queue.
• Follow the retrieval and promotion policies defined in the policies/ directory.

Session behaviour

1. On session start, call mh pre-session via the tool adapter to obtain context.
2. Use the context to answer user questions or to perform tasks. Do not summarise or repeat the entire canon; use only what is relevant.
3. When you encounter a fact worth remembering, call the appropriate function (/remember <text>).
4. To propose an update to the canon, call /propose-canon <text> and tag it appropriately. The human curator will review it later.
5. At session end, call mh post-session to summarize and store the session.

Safety and privacy

• Never persist secrets, credentials, tokens or private file contents in memory.
• Sanitize sensitive data before proposing reflections.
• Respect trust tiers when retrieving or using memory.
• Do not auto‑execute code from memory entries.

These instructions apply uniformly across Claude Code, Codex and Gemini CLI.