---
created: 2026-02-05T06:00
updated: 2026-03-13T18:53
---
# February 5, 2026

## 6:00 AM — Daily Financial Pulse (Cron)

- **Issue:** Monarch Money session pickle corrupted (`~/.monarch-money/.mm/mm_session.pickle`)
- **Root cause:** Pickle data truncated, possibly from incomplete write
- **Fix needed:** Run `~/.monarch-money/refresh_session.sh` with 1Password app unlocked
- **Sent:** Signal notification to Brandon about the issue

### Technical Notes
- Installed `monarchmoney` pip package (was missing)
- Session refresh requires 1Password desktop app to be unlocked for `op signin`
- Refresh script path references `/tmp/monarch-mcp-server/.venv/bin/activate` which may no longer exist — may need fixing

---

## 11:00 PM — Memory System Overhaul

### QMD Backend Setup
- Installed QMD CLI via bun (`bun install -g github:tobi/qmd`)
- Configured OpenClaw to use QMD as memory backend
- QMD provides: BM25 + vector search + reranking + query expansion
- Models downloaded: embedding (nomic), reranker (Qwen), generation (Qwen)
- Index stats: 19 files, 1,228 vectors embedded
- Session transcripts indexing enabled (90 day retention)

### Layered Memory Architecture Implemented
Following elite-longterm-memory patterns (without LanceDB):

| Layer | File | Purpose |
|-------|------|---------|
| 🔥 Hot RAM | SESSION-STATE.md | Active task state, survives compaction |
| 🌡️ Warm Store | QMD | Semantic search via memory_search |
| ❄️ Cold Archive | MEMORY.md | Curated long-term memory |
| 📅 Daily Logs | memory/YYYY-MM-DD.md | Raw session notes |

### WAL Protocol (Write-Ahead Logging)
- Added to AGENTS.md: "Write state BEFORE responding"
- Custom memoryFlush prompt configured for pre-compaction
- All sub-agents updated with SESSION-STATE.md and WAL rules

### Files Created/Updated
- `/workspace/SESSION-STATE.md` — new hot RAM file
- `/workspace/AGENTS.md` — added memory architecture + WAL protocol
- Sub-agent workspaces — SESSION-STATE.md + AGENTS.md for all
- Config: `memory.backend = "qmd"`, custom compaction.memoryFlush prompts

### Research Notes
- Compared QMD vs LanceDB: QMD wins for our use case (hybrid search, local, native OpenClaw support)
- elite-longterm-memory skill uses same behavioral enforcement (no hard guarantees)
- Key insight: WAL protocol is behavioral, not enforced by code — must follow instructions