Percept Ambient

What This Skill Does

Percept Ambient is an advanced contextual intelligence layer for your OpenClaw agent. Unlike standard skills that wait for a command, Percept Ambient operates in the background, continuously processing ambient speech to construct a dynamic knowledge graph. By integrating with percept-listen and percept-summarize, it performs entity extraction and relationship mapping in real-time. It intelligently identifies people, companies, projects, and topics discussed in your environment. The skill uses a hybrid retrieval system, combining FTS5 full-text search with LanceDB vector embeddings (powered by NVIDIA NIM or all-MiniLM-L6-v2) to provide the agent with relevant 'context packets' exactly when they are needed. This allows your agent to maintain high-fidelity memory of meetings, brainstorming sessions, and daily decisions without manual input.

Installation

To begin, ensure you have the prerequisite skills installed. Run the following command in your terminal:

clawhub install openclaw/skills/skills/jarvis563/percept-ambient

Once installed, verify that percept-listen and percept-summarize are active. You can monitor the system, view the entity graph, and configure data retention settings via the local dashboard at http://localhost:8960.

Use Cases

• Project Management: Automatically track project decisions and deadlines discussed during voice meetings.
• Networking: Recall details about people you've met, including their company and past conversations.
• Decision Auditing: Ask the agent about the rationale behind specific past choices that were discussed in passing.
• Passive Knowledge Management: Keep a searchable index of all ambient conversations for later reference without requiring manual documentation.

Example Prompts

1. "What were the key decisions made regarding the Apollo project during today's morning sync?"
2. "Who did I mention I would connect with regarding the marketing strategy last week?"
3. "Based on my recent conversations, what is the current status of the partnership with TechCorp?"

Tips & Limitations

• Privacy: All data is stored locally. Utilize the Privacy settings in the dashboard to set TTL (Time-to-Live) values for auto-purging old transcripts.
• Performance: For high-performance vector search, ensure your environment supports NVIDIA NIM. If not, the system will gracefully fall back to the CPU-based all-MiniLM-L6-v2 model.
• Sensitivity: Because this skill captures ambient audio, be mindful of sensitive or private conversations. You can toggle the recording status via the dashboard or through OpenClaw control commands to ensure privacy in confidential settings.