Long-Term Memory

Last updated: 08.Apr.2026

Current System (as of 08.Apr.2026)

Host: FreeBSD 15.0 bare-metal, ZFS, Bastille jails
Brain: PostgreSQL 17 + pgvector in db jail (10.0.0.3), DB: {agent}_brain
Embeddings: bge-m3 Q4_K_M via llama-server at 10.0.0.5:8080
Chat model: Phi-4-mini-instruct Q4_K_M at 10.0.0.5:8081 (stepping stone; target: Qwen3-14B at 32 GB RAM)
Control Plane: Multi-agent orchestrator running inside clawdie service (port 3100)
- Roles: Orchestrator (80% budget), Sysadmin (10%, daily heartbeat), DBA (5%), Git Admin (5%)
- Main agent id = AGENT_NAME (e.g., clawdie), role = orchestrator
- Identity files: {AGENT_NAME}.md (falls back to SOUL.md), SYSADMIN_AGENT.md, DB_ADMIN_AGENT.md, GIT_ADMIN_AGENT.md
- 15 operational skills in .agent/skills/
Telegram bridge: Routes messages to control plane agents via keyword matching
Dashboard feasibility: Phases A & B green (Vite 8 + React 19 + Tailwind 4 + native Rust bindings all build on FreeBSD 15.0). Phases C & D pending.
No cloud dependency: local llama-cpp replaces former z.ai/GLM/OpenRouter for routine work

About Sam

Name: Samo Blatnik — call him Sam
Location: Near Ljubljana, Slovenia (Europe/Ljubljana, CET/CEST)
Languages: Slovenian (native), Croatian/Serbian, English (good)
Background: IT support (hardware, printers, general troubleshooting)
Telegram ID: 85126311
Dog: Bucko
CV: https://samob.netlify.app/en/

Key Facts

SSH key authentication; ports 80, 443, 22
Identity files: SOUL.md, IDENTITY.md, USER.md backed up in Supabase (backups table)
Memory pipeline active: embeddings → pgvector → semantic retrieval
Deployment goal: bare-metal ISO (clawdie-iso) alongside cloud/VPS option
Default subnet: 10.0.0.0/24 (AGENT_SUBNET_BASE=10.0.0), gateway .1
Jail slots: .2 reserved, .3 db, .4 cms, .5 llama-cpp, .6 git, .101+ workers

Recent Progress (Apr 2026)

[08.04.] Control plane naming cleanup: ceo→orchestrator, company→system, Paperclip→Control Plane across all code + docs
[08.04.] Created CLAWDIE.md as primary orchestrator identity file, wired resolveIdentityFile() fallback
[08.04.] Agent CLI validation: claude, codex, gemini, pi all confirmed working in fresh worker jail
[08.04.] Worker jail template: bash + node24 + git preinstalled, chsh root to bash
[08.04.] Onboarding gate: require at least one agent CLI before proceeding
[08.04.] Archived legacy spec.md (macOS/Docker/NanoClaw references)
[08.04.] Docs: fixed multi-agent subnet defaults to match codebase (10.0.0 first agent)
[06.04.] Phase 4 GUI→shell integration pipeline committed (clawdie-iso)
[06.04.] npm CLI bundling for offline first-boot install (clawdie-iso)

Open Work Items

Dashboard port — Phase C (heavy deps) and D (real Paperclip UI) need FreeBSD testing
IP migration — 10.0.1.x → 10.0.0.x shared IPs across jail configs
Telegram channel integration — replace WhatsApp scaffolding from NanoClaw
FreeBSD jail runtime — replace Docker in container-runner.ts
Supabase memory system — persistent memory layer
Stripe payments — integration
Deploy to ai.clawdie.si — DNS + nginx vhost

Historical Context (Feb–Mar 2026, old system)

The sessions below predate the FreeBSD migration. They ran on a Debian/Hostinger VPS using Docker, OpenClaw v1/v2, and cloud LLMs (GLM, OpenRouter). That infrastructure is retired. These memories are kept for continuity — treat infra references in them as historical, not current.

[24.02.] Major session: 3 skills created, cost optimisation, assistant renamed to Clawdie, Telegram integration
[14.02.] PRD updates v0.3: SSH, Tailscale VPN, user 'clawdie', GLM-5 model configured
[11.02.] Collaborative tmux + Claude Code workflow established
[10.02.] Strapi CMS deployment planning
[09.02.] Memory persistence explained — context files are the persistent layer
[06.02.] Initial setup: Sam configured Clawdie as AI assistant, sudo access, fail2ban, cron

4 KiB Raw Blame History Unescape Escape