# MEMORY.md — Long-Term Memory

## Architecture
- v2 native Ollama tool calling (not text-based parsing)
- Line-based editing tools prevent full-file rewrites
- System prompt must be compact — tools eat most of the 8K context
- Personality files (SOUL, IDENTITY, USER) load per turn, truncated to char limits

## Lessons Learned
- Native tool calling > text-based code generation for reliability
- One action per turn works best for small models
- Line-number tools > find_replace (whitespace matching is unreliable)
- Always write important facts to files — "mental notes" are lost between sessions

## Status (2026-05)
- server-v2.ts: Native tool calling ✓
- Task Runner: Sliding context multi-step ✓
- Browser Automation: Playwright tools ✓
- Skills: Installable via ClawHub ✓
- Email: Multi-account IMAP ✓
- PubMed/Scholar: Search + fulltext ✓
- PPTX: python-pptx ✓
- Weather, Photos, PDF, OCR, Python eval, SQLite: ✓

---
*Use memory_write("memory", ...) to update.*