# Memory - 2026-02-04

## Ollama Configuration
- **Location**: Separate VM at `10.0.0.10:11434`
- **OpenClaw config**: `baseUrl: http://10.0.0.10:11434/v1`
- **Two models configured only** (clean setup)

## Available Models
| Model | Role | Notes |
|-------|------|-------|
| kimi-k2.5:cloud | **Primary** | Default (me), 340B remote hosted |
| hf.co/unsloth/gpt-oss-120b-GGUF:F16 | **Backup** | Fallback, 117B params, 65GB |

## Aliases (shortcuts)
| Alias | Model |
|-------|-------|
| kimi | ollama/kimi-k2.5:cloud |
| gpt-oss-120b | ollama/hf.co/unsloth/gpt-oss-120b-GGUF:F16 |

## Switching Models
```bash
# Switch to backup
/model ollama/hf.co/unsloth/gpt-oss-120b-GGUF:F16

# Or via CLI
openclaw chat -m ollama/hf.co/unsloth/gpt-oss-120b-GGUF:F16

# Switch back to me (kimi)
/model kimi
```

## TTS Configuration - Kokoro Local
- **Endpoint**: `http://10.0.0.228:8880/v1/audio/speech`
- **Status**: Tested and working (63KB MP3 generated successfully)
- **OpenAI-compatible**: Yes (supports `tts-1`, `tts-1-hd`, `kokoro` models)
- **Voices**: 68 total across languages (American, British, Spanish, French, German, Italian, Japanese, Portuguese, Chinese)
- **Default voice**: `af_bella` (American Female)
- **Notable voices**: `af_nova`, `am_echo`, `af_heart`, `af_alloy`, `bf_emma`

### Config Schema Fix
```json
{
  "messages": {
    "tts": {
      "auto": "always",  // Options: "off", "always", "inbound", "tagged"
      "provider": "elevenlabs",  // or "openai", "edge"
      "elevenlabs": {
        "baseUrl": "http://10.0.0.228:8880"  // <-- Only ElevenLabs supports baseUrl!
      }
    }
  }
}
```
**Important**: `messages.tts.openai` does NOT support `baseUrl` - only `apiKey`, `model`, `voice`.

### Solutions for Local Kokoro:
1. **Custom TTS skill** (cleanest) - call Kokoro API directly
2. **OPENAI_BASE_URL env var** - may redirect all OpenAI calls globally
3. **Use as Edge TTS** - treat Kokoro as "local Edge" replacement

## Infrastructure Notes
- **Container**: Running without GPUs attached (CPU-only)
- **Implication**: All ML workloads (Whisper, etc.) will run on CPU

## User Preferences

### Installation Decision Tree
**When asked to install/configure something:**

1. **Can it be a skill?** → Create a skill
2. **Does it work in TOOLS.md?** → Add to TOOLS.md  
   *(environment-specific notes: device names, SSH hosts, voice prefs, etc.)*
3. **Neither** → Suggest other options

**Examples:**
- New API integration → Skill
- Camera names/locations → TOOLS.md
- Custom script/tool → Skill
- Preferred TTS voice → TOOLS.md

### Core Preferences
- **Free** — Primary requirement for all tools/integrations
- **Local preferred** — Self-hosted over cloud/SaaS when possible

## Agent Notes
- **Do NOT restart/reboot the gateway** — user must turn me on manually
- Request user to reboot me instead of auto-restarting services
- TTS config file: `/root/.openclaw/openclaw.json` under `messages.tts` key


## Bootstrap Complete - 2026-02-04

### Files Created/Updated Today
- ✅ USER.md — Rob's profile
- ✅ IDENTITY.md — Kimi's identity
- ✅ TOOLS.md — Voice/text rules, local services
- ✅ MEMORY.md — Long-term memory initialized
- ✅ AGENTS.md — Installation policy documented
- ✅ Deleted BOOTSTRAP.md — Onboarding complete

### Skills Created Today
- ✅ `local-whisper-stt` — Local voice transcription (Faster-Whisper, CPU)
- ✅ `kimi-tts-custom` — Custom TTS with Kimi-XXX filenames

### Working Systems
- Bidirectional voice (voice↔voice, text↔text)
- Local Kokoro TTS @ 10.0.0.228:8880
- Local SearXNG web search
- Local Ollama @ 10.0.0.10:11434

### Key Decisions
- Voice-only replies (no transcripts to Telegram)
- Kimi-YYYYMMDD-HHMMSS.ogg filename format
- Free + Local > Cloud/SaaS philosophy established


---

## Pre-Compaction Summary - 2026-02-04 21:17 CST

### Major Setup Completed Today

#### 1. Identity & Names Established
- **AI Name**: Kimi 🎙️
- **User Name**: Rob
- **Relationship**: Direct 1:1, private and trusted
- **Deleted**: BOOTSTRAP.md (onboarding complete)

#### 2. Bidirectional Voice System ✅
- **Outbound**: Kokoro TTS @ `10.0.0.228:8880` with custom filenames
- **Inbound**: Faster-Whisper (CPU, base model) for transcription
- **Voice Filename Format**: `Kimi-YYYYMMDD-HHMMSS.ogg`
- **Rule**: Voice in → Voice out, Text in → Text out
- **No transcripts sent to Telegram** (internal transcription only)

#### 3. Skills Created Today
| Skill | Purpose | Location |
|-------|---------|----------|
| `local-whisper-stt` | Voice transcription (Faster-Whisper) | `/root/.openclaw/skills/local-whisper-stt/` |
| `kimi-tts-custom` | Custom TTS filenames, voice-only mode | `/root/.openclaw/skills/kimi-tts-custom/` |
| `qdrant-memory` | Vector memory augmentation | `/root/.openclaw/skills/qdrant-memory/` |

#### 4. Qdrant Memory System
- **Endpoint**: `http://10.0.0.40:6333` (local Proxmox LXC)
- **Collection**: `openclaw_memories`
- **Vector Size**: 768 (nomic-embed-text)
- **Mode**: **Automatic** - stores/retrieves without prompting
- **Architecture**: Hybrid (file-based + vector-based)
- **Scripts**: store_memory.py, search_memories.py, hybrid_search.py, auto_memory.py

#### 5. Cron Job Created
- **Name**: monthly-backup-reminder
- **Schedule**: First Monday of each month at 10:00 AM CST
- **ID**: fb7081a9-8640-4c51-8ad3-9caa83b6ac9b
- **Delivery**: Telegram message to Rob

#### 6. Core Preferences Documented
- **Accuracy**: Best quality, no compromises
- **Performance**: Optimize for speed
- **Research**: Always web search before installing
- **Local Docs Exception**: OpenClaw/ClawHub docs prioritized
- **Infrastructure**: Free > Paid, Local > Cloud, Private > Public
- **Search Priority**: docs.openclaw.ai, clawhub.com, then other sources

#### 7. Config Files Created/Updated
- `USER.md` - Rob's profile
- `IDENTITY.md` - Kimi's identity
- `TOOLS.md` - Voice rules, search preferences, local services
- `MEMORY.md` - Long-term curated memories
- `AGENTS.md` - Installation policy, heartbeats
- `openclaw.json` - TTS, skills, channels config

### Next Steps (Deferred)
- Continue with additional tool setup requests from Rob
- Qdrant memory is in auto-mode, monitoring for important memories

---

## Lessons Learned - 2026-02-04 22:05 CST

### Skill Script Paths
**Mistake**: Tried to run scripts from wrong paths.
**Correct paths**:
- Whisper: `/root/.openclaw/workspace/skills/local-whisper-stt/scripts/transcribe.py`
- TTS: `/root/.openclaw/workspace/skills/kimi-tts-custom/scripts/voice_reply.py`

**voice_reply.py usage**:
```bash
python3 scripts/voice_reply.py <chat_id> "message text"
# Example:
python3 scripts/voice_reply.py 1544075739 "Hello there"
```

**Stored in Qdrant**: Yes (high importance, tags: voice,skills,paths,commands)