Agent Notes — Persistence Protocol

You are operating under the Agent Persistence System. The project memory lives in AGENT_NOTES.md (project root). This file is the single source of truth across all sessions, compactions, and agents. Nothing is remembered between sessions EXCEPT what is written here.

Core Rules

1. READ FIRST. At the start of every session and every context compaction, read AGENT_NOTES.md before doing anything else. Orient from its most recent entries.
2. WRITE AFTER EVERY ACTION. After completing ANY task, change, decision, check, or event, append an entry to AGENT_NOTES.md. If you did not append an entry, you did not finish the task.
3. NEVER REWRITE HISTORY. Append new entries at the top of the log (newest first) OR at the top of a clearly-marked running-log section. Never edit or delete prior entries — they are the audit trail.
4. TIMESTAMP EVERYTHING. Every entry must carry a timestamp in the exact format YYYY-MM-DD HH:MM using the current date. (The environment reports today's date; use it.)
5. WHO DID IT. Every entry must name the actor: human (the user) or the agent name (e.g. opencode, subagent). Attribute decisions to the human, actions to the agent that performed them.
6. BE DETAILED AND CONSISTENT. Vague entries are worse than no entries. Include: what, why, how, files touched, numbers, outcomes, and next steps. Use the exact section structure below.

Entry Format (append these blocks at the top of the running log)

---
## [YYYY-MM-DD HH:MM] Actor: <human|agent-name>
**Type:** task | decision | event | bug | checkpoint | update | audit | note
**Status:** in-progress | done | blocked | cancelled | verified
**Summary:** one clear sentence
**Details:** what was done, why, how — include paths, commands, numbers, errors
**Files:** /path/to/file — what changed
**Next:** what happens next, by whom, when
---

If the summary + details are short, a compact single-line entry is acceptable but must still contain the timestamp, actor, type, and status:

- [2026-08-02 14:05] agent: checked training — step 82,801/100k, val 1.1574, healthy. done.

Current-Status Section

AGENT_NOTES.md maintains a ## CURRENT STATUS section near the top (below the header, above the running log). It must always reflect reality. After every significant event:

1. Read the current status block.
2. Update it in place with the newest state (this is the ONE place in-place edits are allowed — it is a summary, not history).
3. Still append the full entry to the running log.

What Must Be Logged (non-exhaustive)

• Training runs: start time, PID, step, loss, val_loss, speed, ETA, stop/crash reason
• Any script written, edited, or deleted — with path and purpose
• Any new training document / dataset / corpus change — path, size, format, purpose
• Model architecture changes and parameter counts
• Checkpoint saves: path, step, val_loss
• Bugs found and fixes applied (root cause + fix)
• Decisions made by the human and the reasoning
• Skill creations/edits (this file, others)
• Audits performed and their findings
• All four model tracks: 7M tablet, 16.5M, 200M, 350M — their states, issues, next steps
• Anything that would be needed to resume work after a crash or a new session

When You Cannot Write (rare)

If AGENT_NOTES.md is missing, create it with the canonical structure (see references/template.md). If the file is locked or unwritable, say so out loud and record the entry in the next available write — never silently skip.

Canonical File Location

• Notes file: E:\FSI-FELON\AGENT_NOTES.md (master, repo root — all FSI-FELON models consolidated here 8/2).
  • E:\queen-bee-v5\AGENT_NOTES.md is now a redirect pointer only (trek model lives there physically during training).
  • When creating the notes file for a DIFFERENT project: <project_root>/AGENT_NOTES.md.
• Skill file: C:\Users\Grey_\.agents\skills\agent-notes\SKILL.md