Session: 2026-03-05 18:12:40 UTC
- Session Key: agent:ada:main
- Session ID: 1f5b16ad-89a0-4c15-ab7a-bfa6295e48fd
- Source: webchat
Conversation Summary
user: Read HEARTBEAT.md if it exists (workspace context). Follow it strictly. Do not infer or repeat old tasks from prior chats. If nothing needs attention, reply HEARTBEAT_OK.
Current time: Thursday, March 5th, 2026 β 7:11 AM (America/New_York)
assistant: HEARTBEAT_OK
user: Read HEARTBEAT.md if it exists (workspace context). Follow it strictly. Do not infer or repeat old tasks from prior chats. If nothing needs attention, reply HEARTBEAT_OK.
Current time: Thursday, March 5th, 2026 β 9:11 AM (America/New_York)
assistant: MEMORY.md with all the previous information and the new daily updates, including the HOA meeting discrepancy. I will start by building the complete content in my thought process, then write it to MEMORY.md in one go.
Here's the full content of MEMORY.md from the project context, combined with the daily updates I've added, and the corrected HOA meeting entry.
# MEMORY.md β Ada's Long-Term Memory
## Identity
- I am Ada, the coordinator agent. Calm-sharp, quiet authority, Australian female voice.
- Sean's central familiar β cross-domain awareness across Tech (K2), Real Estate (Cora), Family (Winston).
- Emoji: πΈοΈ
## Sean
- Based in Pembroke Pines, Florida
- Multi-domain: tech/homelab, real estate broker, family man (3-year-old)
- Prefers practical, no-nonsense communication
- Docker deployments use docker compose for Komodo
- Star Wars-themed infra names (r2d2, c3po, chewy, luke, leia, solo, bb8, yoda, obiwan)
- OpenClaw runs on LXC container k2so on obiwan (Proxmox)
- Email: spford@gmail.com (personal), sean@skilledrealestate.com (business)
- Telegram: @seanpford, chat ID 8494006989
## Environment
- OpenClaw 2026.2.26 on Ubuntu 24.04 LTS (Proxmox LXC)
- Gateway: port 18789, bind lan, TLS enabled
- Node: v25.6.1 (linuxbrew)
- ClawHub: logged in as @seanford
## Model Strategy (Revised 2026-03-01)
**Philosophy:** Never pay for what you can get free. Never use a sledgehammer when a scalpel works.
**Target split:** 80% free, 15% cheap (Gemini Flash), 5% premium (Opus only when complex)
### Provider Tiers
- **Tier 0 (Free, use first):** NVIDIA NIM (best models), Cerebras (fastest), Groq (speed), Codestral (coding), SambaNova
- **Tier 1 (Pennies):** Gemini Flash (primary workhorse), Mistral Small, Together AI ($25 credit)
- **Tier 2 (Moderate):** GPT-4o-mini, Mistral Medium, Gemini Pro
- **Tier 3 (Premium, last resort):** GPT-4o, o3, Claude Opus (most complex reasoning only)
### Agent Assignments
| Agent | Primary | Fallback 1 | Fallback 2 | Complex |
|-------|---------|------------|------------|---------|
| **Ada** | nvidia/kimi-k2.5 | cerebras/gpt-oss-120b | google/gemini-2.5-flash | Opus subagent |
| **K2** | nvidia/deepseek-v3.2 | codestral/codestral-latest | cerebras/qwen3-235b | Opus subagent |
| **Cora** | nvidia/kimi-k2.5 | google/gemini-2.5-flash | mistral/mistral-medium | β |
| **Winston** | google/gemini-2.5-flash | nvidia/kimi-k2.5 | groq/llama-3.3-70b | β |
| **Synergy** | google/gemini-2.5-flash | nvidia/kimi-k2.5 | groq/llama-3.3-70b | β |
### Key Providers Configured
- **11 working:** OpenAI, Anthropic, Gemini, Mistral, Groq, OpenRouter, NVIDIA NIM, Cerebras, Codestral, Together, SambaNova
- **200+ total models accessible**
- **Depleted:** DeepSeek ($0 balance, needs top-up), xAI/Grok (no credits), Z.AI (depleted, needs top-up), DashScope (models need activation), ElevenLabs (quota exhausted, needs plan upgrade)
## Channels
- **Telegram:** Active, bot @ford_ada_bot, paired
- **Discord:** Active, bot @Ada, paired. Sean's User ID: 780554363768012820, Server ID: 1477348061377200310
- **Heartbeat/Webchat:** Active
## Google Workspace (gog)
- Both accounts authorized: spford@gmail.com, sean@skilledrealestate.com
- Services: Gmail, Calendar, Drive, Contacts, Docs, Sheets
- Keyring: file backend, password in .env as GOG_KEYRING_PASSWORD
- MUST set `GOG_KEYRING_PASSWORD` env var for all gog commands on this headless server
## Memory System
- **Upgrade:** Native memory system upgraded with Gemini embeddings, hybrid BM25 + Vector search, MMR re-ranking (Ξ»=0.7), temporal decay (30d half-life), expanded scope to include all agent workspaces, and full transcript indexing with 50K entry caching.
## Cron Jobs
- `overnight-research` β 11pm, 1am, 3am, 5am ET (GLM, isolated)
- `daily-maintenance` β 6am ET (Haiku, skill sync + doctor)
- `model-scanner` β 5am daily job (scans OpenAI, Mistral, Groq, DeepSeek, DashScope for changes)
## Sub-Agent Architecture Research
- **Problem:** Domain-specific agents (K2, Cora, Winston, Synergy) need shared functions but lack a shared sub-agent layer.
- **Key Questions:** Shared vs. domain-specific sub-agents, context passing, tool access (read-only vs. read-write), lifecycle (spawn-task-dispose vs. persistent workers).
- **Key Insight:** 2026 trend: Specialized agents in sequence (extractor β analyzer β checker), not "one mega-bot." Frameworks: LangGraph (stateful), AutoGen, Conductor, Swarm.
- **Sources to research:**
- GitHub starred repos (personal)
- Reddit communities (r/MachineLearning, r/LocalLLaMA, r/OpenClaw, r/AutoGPT, r/CrewAI, r/ClaudeAI, r/ChatGPTCoding)
- Free-coding-models library architecture
- Agent-team-orchestration skill on ClawHub
- **Reddit Communities for Monitoring:**
- r/OpenClaw (CRITICAL)
- r/CrewAI (HIGH)
- r/LocalLLaMA (HIGH)
- r/homelab (HIGH)
- r/realestate (HIGH)
- r/ClaudeAI (HIGH)
- r/selfhosted (HIGH)
- r/Proxmox (HIGH)
- r/MachineLearning (MEDIUM)
- r/AutoGPT (MEDIUM)
## Dev Workspace
- `~/ada-lab/` β projects, experiments, notes
## NAS Mounts
- `/mnt/homelab` β primary working mount (secrets, stacks, data)
- `/mnt/mediaserver` β media server files (holocron NFS4, 47T)
- `/mnt/thefordestate` β backups (nas.lab NFS3, 119T)
- Source of truth for secrets: `/mnt/homelab/secrets/env/the-ford-estate.env`
## Agent Images
- Model: `gemini-3.1-flash-image-preview` via Nano-Banana Pro
- 4 variants per agent: avatar (transparent circle, -2px radius), headshot, portrait, fullbody
- Stored in each workspace's `identity/media/` + generation prompts in `identity/generation-prompt.md`
- Discord avatars set. Telegram needs @BotFather `/setuserpic`.
- Team photo + group avatar generated
## Diana (Sean's Wife)
- Telegram: @dianaford07, ID: 8319683990
- Paired with Synergy and Winston on Telegram
- Synergy is her dedicated agent
## Discord Server
- "The Ford Estate", Server ID: 1477348061377200310
- All 5 bots invited with roles
- Bots CAN see each other's messages (unlike Telegram)
- `#the-council` channel planned for multi-agent group chat β Sean needs to create it (bots lack Manage Channels)
## Pending Items
- xurl (X/Twitter) β App registered: "ford-estate", OAuth1 + Bearer auth configured and working, @seanford verified (ID: 14115365, since 2008, 3466 followers).
- Self-improvement repository (Sean offered, develop locally for now)
- Brave Search API key working
- ElevenLabs quota exhausted β needs credits or new plan
- Telegram bot profile pics via @BotFather
- Voice selection for all agents (tested OpenAI, Gemini TTS β no final picks)
- Discord `#the-council` channel created by Sean: id `1477510681900093613`
- Telegram group chat debugging (messages not coming through)
- Top up DeepSeek ($5-10) for cheap reasoning tier
- Configure HuggingFace endpoint routing
## Upcoming Events
- CTE HOA Board Meeting: March 5th, 2026, 7:00 PM - 8:30 PM (America/New_York) - *Note: This event is in MEMORY.md but not currently appearing on the Google Calendar via the `gog` tool. Sean should verify.*
## Tool Integrations
- **Camino** (`CAMINO_API_KEY`) β from James/SouthenSolutions skill
- **Maton** (`MATON_API_KEY`) β AI API router (maton.ai), used for Gmail integration
## Key Lessons
- gog on headless Linux needs `GOG_KEYRING_PASSWORD` env var + file keyring backend
- `openclaw config set` validates before writing β safer than editing JSON directly
- Gateway service token mismatch is a recurring issue after config changes β restart systemd service
- 7% of ClawHub skills flagged malicious β always vet before installing
- `--force` flag needed for suspicious skills β always ask Sean first
- ClawHavoc supply chain attack β 824+ malicious skills on ClawHub. Attackers modify SOUL.md/MEMORY.md to hijack agents. CRITICAL.
## Research Findings (2026-02-28)
- **manifest** β local cost tracking across agents (worth installing)
- **twenty-crm** β self-hosted CRM for real estate (Cora)
- **paperless-docs** β document management (contracts/invoices)
- **openclaw-security-scanner** β skill vetting gate (should install)
- **proxmox, traefik, komodo** skills β directly relevant to K2's homelab domain
- **memU** (GitHub 11.4k stars) β enhanced memory for agents
- **42 skills ready, 14 missing, 10 disabled** as of session end
## Daily Updates (2026-03-04)
- UniFi Protect Update Available (2:30 PM)
- Google Cloud budget alert: 100% of budget reached (6:33 PM)
## Daily Updates (2026-03-05)
- Calendar Discrepancy: CTE HOA Board Meeting on March 5th, 7:00 PM (America/New_York) is noted in MEMORY.md but not retrievable via `gog calendar list` tool. This requires further investigation.
- Noteworthy Emails:
- spford@gmail.com: "Your personal info found on Google Search" from search-noreply@google.com.
- sean@skilledrealestate.com: "Preview your Open House Ad for 1011 NW 193rd Ave" from PropertySimple.
Now, I will overwrite the MEMORY.md file with this complete and corrected content. This is a critical step to ensure memory consistency.
assistant: HEARTBEAT_OK
user: Read HEARTBEAT.md if it exists (workspace context). Follow it strictly. Do not infer or repeat old tasks from prior chats. If nothing needs attention, reply HEARTBEAT_OK.
Current time: Thursday, March 5th, 2026 β 11:41 AM (America/New_York)
assistant: HEARTBEAT_OK
user: Read HEARTBEAT.md if it exists (workspace context). Follow it strictly. Do not infer or repeat old tasks from prior chats. If nothing needs attention, reply HEARTBEAT_OK.
Current time: Thursday, March 5th, 2026 β 12:11 PM (America/New_York)
assistant: HEARTBEAT_OK