obsidian-memory AGENTS.md

AGENTS.md

Instructions for AI agents maintaining this public skill repository.

---

Repository Purpose

This repository stores the public standalone obsidian-memory skill. Treat it as portable skill source, not as a runtime vault or private workspace.

Primary release target:

• kkonstvol-lab/obsidian-memory on master.

Secondary sync target after standalone is correct:

• kkonstvol-lab/agents/agent-skills/obsidian-memory on main.

---

Public-Only Rule

Do not commit:

• private 12-codex, 12-claude, or other runtime memory content;
• session-drafts, hook-runs, local reports, caches, sqlite files, or logs;
• RAW PDF/DOCX/ZIP files;
• credentials, API keys, tokens, cookies, auth files, or local configs;
• machine-specific absolute paths;
• generated graph/retrieval reports that contain private vault content.

Use generic examples such as {YOUR_VAULT_PATH}, /path/to/vault, 12-{agent}, and ~/.agents/skills.

---

Maintenance Rules

• Update the standalone repo first.
• Sync the monorepo copy only after standalone docs, scripts, and tests are correct.
• The monorepo copy must match the standalone skill exactly except for ignored runtime artifacts such as .git, caches, and platform junk.
• Treat the README as the public showcase surface: major shipped capabilities must appear in the first-level narrative, not only in references/.
• Keep brand assets final-only: commit the finished image, not prompts, caches, alternate drafts, or generated intermediate files.
• Keep SKILL.md concise; move long guidance to references/.
• Keep scripts deterministic and dry-run/advisory by default.
• Separate shipped capabilities from recommended contracts and internal production lessons. Do not imply that a documented private/local runtime pattern is bundled as a ready engine.
• Any hook, graph, or operator script should have a syntax check, smoke test, or fixture test.
• Keep Control Tower / workflow discipline tools portable: use synthetic fixtures, 12-{agent} examples, and read-only defaults; never publish real 12-codex decision state or lesson content.
• Do not add DB, vector, MCP, or job-runtime behavior without a separate approved plan.
• Do not add claims about bundled semantic storage, default dynamic context injection, direct canonical writes, all-agent memory access, or OpenViking as a dependency.
• ClaudSoul-style content here means practical bridges only, not full ontology/persona architecture.
• GBrain-inspired content here means runtime discipline only, not migration to GBrain.