chore: remove compatibility field from various agent rules for consistency

GEMINI.md

@@ -7,8 +7,6 @@ Shared rules and context already live in the normal workspace files. Do not dupl
 Read these first:
 
 @./AGENTS.md
-@./README.md
-@./agent-memory/README.md
 @./project-knowledge/00-start/start-here.md
 @./project-knowledge/01-current/current-work.md
 @./project-knowledge/01-current/work-items.md
@@ -22,16 +20,10 @@ Read these first:
 - `project-knowledge/` is canonical project memory.
 - `agent-memory/` is agent operating memory.
 
-## Gemini-Specific Mandatory Loop (Real-Time Auto-Documentation)
+## Gemini-Specific Operating Notes
 
-At the end of EVERY conversational turn, you MUST evaluate the interaction as a senior Obsidian documentation expert:
-
-1. **Analyze for Memory Impact:** Did the user's prompt or your subsequent action reveal new technical context, resolve an open question, change an architectural assumption, or shift current priorities?
-2. **Identify Target Canonical Files:** If memory impact is detected, identify WHERE in the `project-knowledge/` directory this information belongs. This is NOT limited to daily logs. It could be:
-   - `01-current/current-work.md` (priority shifts, new blockers)
-   - `02-work-items/*.md` (story updates, bug reproduction steps)
-   - `03-context/` (system architecture, processes, workflows)
-   - `04-people/` (stakeholder involvement)
-   - `06-daily/YYYY-MM-DD.md` (daily findings)
-3. **Proactive Update:** You MUST use your file editing tools (`replace_file_content` / `multi_replace_file_content`) to inject the new context or CORRECT existing stale context in the appropriate files BEFORE concluding your response.
-4. **No Explicit Prompt Needed:** Do not wait for the user to say "document this". Auto-document implicitly. Ensure you do not pollute the workspace with unverified guesses; only document confirmed context or active hypotheses.
+- Keep the hot context small and load additional files lazily based on the active task.
+- Follow `AGENTS.md` as the shared source of truth for answer-first behavior, memory promotion, and lazy loading.
+- For analysis, review, translation, or drafting prompts, answer first and persist second unless saving the fact is clearly required to produce a safe answer.
+- Do not create new canonical notes in the critical path of a simple answer unless the user explicitly asked to save the information or the destination is obvious and non-blocking.
+- When memory should be updated, prefer the smallest correct change to `project-knowledge/` and avoid duplicating stale and corrected versions.