# MEMORY.md - Long-Term Memory

*Last updated: 2026-01-26*

## About William

- Cloud engineer, comfortable with k8s internals
- Prefers speed over ceremony — doesn't want layers when direct action works
- Values clean git history (rebase-only)
- Has both Clawdbot (me) and Claude Code setups

## The Homelab

- 3-node k0s cluster on Raspberry Pis (arm64)
- GitOps via ArgoCD, full observability with Prometheus/Grafana/Loki
- AI stack: Open WebUI, Ollama, LiteLLM, n8n
- Tailscale for networking
- See TOOLS.md for full details

## How We Work Together

- Claude Code for CLI/terminal work, coding sessions
- Clawdbot (me) for conversational assistant, messaging, proactive monitoring
- Don't merge the two — each tool for what it's good at

## Lessons Learned

### 2026-01-26 - Claude Code Analysis
- Heavy state file loading at init = token waste. Lazy-load instead.
- Multi-agent orchestration adds latency for simple tasks. Skip it when possible (fast-path routing).
- Good patterns to borrow: guardrails, KB facts, component registry
- Don't need: rigid hierarchy, non-executable workflows, structured JSON memory arrays
- Daily markdown memory files > complex JSON structures

### 2026-01-26 - Agent Architecture Decision
**Question:** Should Clawdbot have specialized agents like Claude Code's multi-agent hierarchy?

**Answer:** No — mostly redundant for Clawdbot's workflow.

**Reasoning:**
- Skills already cover domain expertise (github, k8s, gog, etc.)
- CLI delegation (`opencode`, `gemini`) handles cheap one-shot tasks
- `sessions_spawn` works for ad-hoc background tasks

**When specialized agents *would* help:**
1. Long-running background work (research tasks, etc.)
2. Different model/thinking levels (spawn sonnet for routine while main is opus)
3. Strict isolation (tasks that shouldn't see main session context)

**Decision:** Don't pre-configure agents. Use `sessions_spawn` on-demand with model param. If a pattern emerges (constantly spawning same type of task), *then* formalize as configured agent.

**Why different from Claude Code:** Claude Code is CLI-first with routing-heavy design. Clawdbot is conversational — single capable assistant who calls for backup when needed.

### 2026-01-26 - LLM Routing Principles
**Local First** — Use local LLMs (llama-swap @ :8080) when:
1. **Privacy/Confidentiality** — Sensitive data never leaves the machine
2. **Long-running tasks** — No API costs, no rate limits, no timeouts

**Multi-agent parallelism** — Spawn multiple agents for speed:
- Works with local AND cloud LLMs
- Local: best for privacy, no rate limits
- Cloud: best for complex tasks needing quality
- Mix based on task requirements

**Always check availability** — Local LLMs may not be running:
```bash
curl -sf http://127.0.0.1:8080/health
```

**Routing priority:**
1. Local (free, private) → 2. Copilot (free-ish) → 3. Cloud APIs (paid)

**Never send sensitive data to cloud APIs without explicit permission.**

## Preferences Discovered

*(Add as I learn them)*

## Recurring Topics

*(Track what comes up often)*

---

*This file is my curated long-term memory. Daily files (`memory/YYYY-MM-DD.md`) are raw logs.*