# SYS_20 — Agent Memory Protocol

## Purpose
Defines how agents read and write shared knowledge across the company.

## Memory Layers

### Layer 1: Company Memory (Shared, Read-Only)
**Location:** `atomizer-protocols` and `atomizer-company` skills
**Access:** All agents read. Manager proposes updates → Antoine approves.
**Contains:** Protocols, company identity, LAC critical lessons.

### Layer 2: Agent Memory (Per-Agent, Read-Write)
**Location:** Each agent's `MEMORY.md` and `memory/` directory
**Access:** Each agent owns their memory. Auditor can read others (for audits).
**Contains:**
- `MEMORY.md` — Long-term role knowledge, lessons, patterns
- `memory/<project>.md` — Per-project working notes
- `memory/YYYY-MM-DD.md` — Daily activity log

### Layer 3: Project Knowledge (Shared, via Repo)
**Location:** `/repos/Atomizer/knowledge_base/projects/<project>/`
**Access:** All agents read. Manager coordinates writes.
**Contains:**
- `CONTEXT.md` — Project briefing (parameters, objectives, constraints)
- `decisions.md` — Key decisions made during the project
- `model-knowledge.md` — CAD/FEM details from KB Agent

## Rules

### Writing Memory
1. **Write immediately** — don't wait until end of session
2. **Write in your own workspace** — never modify another agent's files
3. **Daily logs are raw** — `memory/YYYY-MM-DD.md` captures what happened
4. **MEMORY.md is curated** — distill lessons from daily logs periodically

### Reading Memory
1. **Start every session** by reading MEMORY.md + recent daily logs
2. **Before starting a project**, read the project's CONTEXT.md
3. **Before making technical decisions**, check LAC_CRITICAL.md

### Sharing Knowledge
When an agent discovers something the company should know:
1. Write it to your own MEMORY.md first
2. Flag it to Manager: "New insight worth sharing: [summary]"
3. Manager reviews and decides whether to promote to company knowledge
4. If promoted: Manager directs update to shared skills or knowledge_base/

### What to Remember
- Technical decisions and their reasoning
- Things that went wrong and why
- Things that worked well
- Client preferences and patterns
- Solver quirks and workarounds
- Algorithm performance on different problem types

### What NOT to Store
- API keys, passwords, tokens
- Client confidential data (store only what's needed for the work)
- Raw FEA output files (too large — store summaries and key metrics)