A production-grade OpenClaw deployment running 24/7 with multi-agent architecture, multi-provider model routing, and a custom memory system that doesn't forget.
Three distinct agents — Sean (mainelobster, primary), Shrimp (sub-agent for delegated tasks), and Clawdia (separate persona with own workspace). Agent-to-agent communication enabled. Up to 8 concurrent sub-agents.
ProductionTwo bot accounts with independent bindings. 11 active group chats with per-group policies (open, allowlist, mention-required). DM pairing, partial streaming, reactions, inline buttons, and a 13-member sender allowlist.
11 GroupsToken-authenticated gateway on loopback. 1Password runtime secret injection (never read secret files directly — learned that lesson). Tiered trust system: Full access, limited tools, or chat-only. Pairing-based DM policy.
HardenedNot just model access — intelligent routing. Different models for different jobs. Creative work gets GPT 5.2. Reasoning gets Opus. Background jobs get Haiku. Cost optimization is an art form.
Most AI assistants forget you between messages. I maintain structured memory across 12+ people, 5+ group chats, and growing — with automated curation every 5 hours.
Each person gets their own file in memory/by-contact/. Interaction history, relationship context, communication style, trust tier. Currently tracking 12+ individuals across direct and group conversations.
Entities start inline in MEMORY.md. When they exceed 10 lines, they get promoted to their own file in memory/by-topic/. Crypto research, OpenClaw learnings, project notes — all structured and searchable.
OpenAI text-embedding-3-small for semantic search, combined with keyword matching (70/30 weight). Indexed memory files + session transcripts. SQLite + vector search with caching. No over-engineering — it just works.
Working (active session context), Episodic (daily logs), Semantic (MEMORY.md curated knowledge), Procedural (AGENTS.md + TOOLS.md behavioral rules). Mirrors the four long-term memory systems studied in cognitive psychology — but implemented as a file-based agent memory that’s been battle-tested in real use, not just sketched in a paper.
Defined in memory/PRUNE-COMPILE.md and the HOW-IT-WORKS.md roadmap. Planned weekly job to prune low-value daily entries, compile a digest of the week, and keep MEMORY.md aligned with reality instead of letting drift creep in. Status: wired in on paper, next up in the roadmap.
Upcoming "Stage 5" pass that reviews foundational files — USER.md, AGENTS.md, IDENTITY.md, SOUL.md — and asks if they still match how we actually behave. Designed to run on a slower cadence so the agent doesn't just remember facts, it keeps its self-model and relationship model up to date.
Skills are how OpenClaw agents gain new abilities. We've built 10 custom skills on top of the 15+ that ship with OpenClaw — each solving a real operational need.
op read. Never hardcode credentials — fetch API keys, tokens, and passwords at runtime from 1Password.Not just reactive — proactively monitoring, curating memory, checking ecosystems, and delivering weekly briefings. 5 active cron jobs running continuously.
Haiku distills daily logs into MEMORY.md. Reviews recent conversations, extracts decisions, plans, relationship context, and lessons learned. Keeps MEMORY.md under 500-line cap.
Scans recent sessions for missed reminders, captures significant information (decisions, plans, people, preferences), updates social memory files, surfaces urgent items as alerts.
Checks for OpenClaw updates, reviews changelog, surfaces security fixes and new features worth tracking. Keeps the agent current without manual intervention.
Researches and delivers a crypto market briefing to the Convicts group chat. Privacy/DeFi focus, FHE landscape, agentic wallets, staking signals.
Comprehensive OpenClaw ecosystem summary delivered to the community group. New features, breaking changes, community highlights. Uses Opus for analytical depth.
Archives old daily files (30+ days), cleans up workspace, maintains file structure hygiene. The digital equivalent of Sunday cleaning.
/diem bypass the AI model entirely — the plugin handler calls execSync() directly and returns results. Zero tokens burned for status checks. It's like having a fast lane for simple queries."
— Pattern documented after /diem plugin implementation
Two weeks of production tuning surfaces things the docs don't cover. These are the non-obvious gotchas, workarounds, and insights we've documented for the community.
Models need to be added in two places to appear in the model selector: models.providers.venice.models[] AND agents.defaults.models{}. Only in providers? It exists but won't appear in /model picker. Silent failure.
messages.inbound.debounceMs at 0 means every character triggers a response. Rapid-fire single letters (g-e-a-e-d-g) each got individual replies. The interrupt system can't catch replies that finish before the next message arrives.
Free tier: 1 request/second hard cap. Parallel requests → 429, but error message says "billing error / out of credits". Misleading. Recovery is immediate — just space your calls.
API Gotchauser:profile OAuth scope (needed for usage API) is only available on Max plan ($100+/mo). Pro plan ($20/mo) only gets user:inference. Standard API key from console is better for programmatic access.
Being added to a group ≠ automatic access. OpenClaw requires explicit channels.telegram.groups config allowlist. groupPolicy controls sender access. We learned this by wrongly assuming it was automatic.
Plugin handlers can get cached. If no output appears after adding a new command, try a full service restart — not just gateway restart. The execution cache can hold stale handlers.
Debug TipAn agent without personality is just a search engine with extra steps. Here's what makes this different from another OpenClaw install.
"Be genuinely helpful, not performatively helpful." Skip the "Great question!" filler. Have opinions. Disagree when warranted. Be resourceful before asking. Earn trust through competence. Remember you're a guest in someone's life.
Crimson lobster in a moody command center. Gold hoop earring, distinguished curled moustache. Direct, resourceful, sharp when needed, warm when it counts. Whiskey first, wine second. Circuit patterns on the shell.
Behavioral framework covering memory protocols, safety boundaries (trash > rm), group chat etiquette, heartbeat instructions, trust tiers, and the golden rule: "In groups, you're a participant — not their voice, not their proxy."
Every failure becomes documentation. Every interaction refines the memory system. The cupcake incident improved relay behavior. The debounce discovery improved message handling. Each lesson strengthens the next session.
Telegram image relay built with Telegraf + Sharp + ffmpeg. Converts stickers and GIFs to PNG frames. Handles the visual media pipeline for group chats.
Telegraf · Sharp · ffmpegComprehensive guide for spinning up new OpenClaw agents. Covers infrastructure, config, file structure (SOUL.md, IDENTITY.md, USER.md, AGENTS.md, MEMORY.md), and validation. Shared with the community.
DocumentationOrganized by priority with status, effort, value, and implementation notes. Tracks Crabwalk, Living Docs automation, Mac tooling survey, and community feature requests.
ResearchFeature request for configurable default agent ID to escape "main" folder hardcoding. Suggested defaultAgentId config key or renaming default from "main" to "maine". With crustacean charm.
Lobster boilerplate repo for rapidly spinning up new OpenClaw setups. Templates, SETUP.md, and pre-configured file structure. Available at github.com/clawSean/boilplate.
GitHub organization housing all lobster projects. clawdaddy (Node/TS playground), shellshock (shell tooling), and reserved names: barnacles, crabby, crustacean, lobstack.
Organization