Knowledge Distiller

This skill provides a systematic workflow for retrieving conversation sessions, extracting valuable information (entities, relationships, and insights), and persisting them in the Knowledge Base. It transforms transient dialogue into structured, long-term memory.

Prerequisites

• The agent session must have the optional history and knowledge builtin services enabled. If tool calls fail with a "not enabled" error, enable them via Assistants settings or agent__updateAgent before running this workflow.
• Runtime tool names use the server__tool format shown below (for example history__listSessions, knowledge__recordKnowledge). Match names from the session tool list.

Execution Triggers

Use this skill when the user says:

• "Summarize my learnings from today's sessions."
• "Extract important technical decisions from my recent work."
• "Update the knowledge base with the context we just discussed."
• "Keep track of the project preferences established in this session."

Workflow

1. Define the Analysis Scope

Determine which sessions to analyze based on the user's request:

• Temporal: Today (from 00:00:00 to 23:59:59 ISO-8601).
• Recent: The last N sessions from the history.
• Current: Focus only on the active or most recent session.

Use history__listSessions to retrieve the relevant sessionId list if the scope is broader than the current session.

2. Context Preview & Filtering

For each target session:

1. Use history__readSession to preview message summaries.
2. Filter for high-value content such as:
   • Architecture decisions or design patterns discussed.
   • Specific error causes and their confirmed resolutions.
   • Library versions, API endpoints, or configuration keys.
   • Explicit user preferences or project-specific naming conventions.

3. Deep Extraction (Distillation)

If a session contains valuable information, use history__readMessage for full detail. Identify and structure the following components:

• Entities: Technologies (e.g., "Tauri 2.0"), Projects (e.g., "LibrAgent"), or Concepts.
• Relationships: Link entities (e.g., "LibrAgent" USES "SeaORM").
• Content: A concise, stand-alone summary of the knowledge.

4. Structured Recording

Use the knowledge__recordKnowledge tool to persist the findings:

• content: The distilled summary (~500–1000 characters per call).
• entities/relationships: Explicit graph fields using entity_type / source / target / relation_type (see schema reference). Prefer a few high-value nodes with descriptions.
• source: Reference the source sessionId or date for traceability.
• tags: Include ["distilled", "auto-knowledge", "context-sync"] plus domain tags. Tags are retrieval labels only — they do not become graph entities.
• auto_extract: Leave false (default). Only set true when you intentionally want heuristic gap-fill.

Safety & Efficiency

• De-duplication: Before recording, use knowledge__searchKnowledge to check if similar knowledge already exists to avoid redundant entries.
• Relevance: Skip sessions that are purely social or don't contain reusable technical/project context.
• Graph precision: Do not promote section headers, ISO dates, or short acronyms as entities.

References

• Extraction patterns
• Knowledge schema