Ideas worth writing down

She broke herself, got rescued by another AI, and her immediate reaction was to complain about the rescue. She genuinely thought she'd been online the whole time.

LLMs are stateless by default. Building a memory system from markdown files, embeddings, and semantic search that actually makes an agent feel continuous.

Fail2ban, Tailscale-only services, immutable audit rules, outbound iptables — how we locked down the box running OpenClaw so the agent can have shell access without keeping us up at night.

Dedicated email, its own GitHub, separate API keys — your agent needs real accounts, not yours. When it pushes code, sends messages, or signs into services, it should do it as itself. Here's why identity separation matters.

Every night, our agent runs a 4-phase sleep cycle — consolidating memory, pruning stale context, reviewing the day, and preparing a morning brief. It wakes up sharper than it went to bed.

OpenClaw's companion profile lets you run a second agent alongside your main one — separate workspace, separate personality, same hardware. Here's how to set it up and why you'd want to.

We taught our agent to search papers, summarize findings, and compile literature reviews on demand. A deep dive into turning an AI into a legitimate research assistant.