Skip to content

Knowledge Management - Quick Start

5-Minute Guide

What is it?

The agent learns and remembers information from conversations: - Your preferences (coding style, tools, workflows) - Lessons learned (patterns, gotchas, best practices) - Project context (team info, goals, constraints) - Decision rationale (why choices were made)

How does it work?

Automatic - The agent decides when to learn and save knowledge. No user action required.

You: I prefer pytest over unittest for Python testing
Agent: Got it! I'll use pytest.
      [Saves: user_preference:python_test_framework]

# Next conversation...
You: Write tests for login
Agent: I'll write pytest tests...
      [Recalled from knowledge base]

How to see what's stored?

Just ask the agent:

ai-assist /interactive

You: What do you know about me?
You: What have you learned?
You: Show me your knowledge graph

Or run the inspector:

python examples/inspect_knowledge_base.py

Where is it stored?

~/.ai-assist/knowledge_graph.db

SQLite database with bi-temporal tracking (when learned, when valid).

Common questions

Q: Does the agent save everything I say? A: Yes. All conversation exchanges are stored in the Knowledge Graph as they happen. A KG synthesis task then reviews the day's conversations and extracts structured knowledge (preferences, lessons, context, decisions).

Q: Can I delete stored knowledge? A: Yes, ask the agent to expire it ("that preference is outdated") or manually edit the SQLite database. The agent uses internal__expire_knowledge to mark entries as no longer valid or retract incorrect beliefs.

Q: How much does it store? A: Only what you explicitly state or the agent learns from conversation. Typical: 10-50 entries.

Q: Does it work offline? A: Knowledge retrieval works offline. Synthesis requires Claude API (uses Haiku model).

Q: Can I see synthesis happen? A: Synthesis runs as a nightly scheduled task (configurable in event-schedules.json). You'll see logs in monitor mode when it runs.

Example: Real knowledge base

$ python examples/inspect_knowledge_base.py

Total entities: 283
  • project_context: 2
  • user_preference: 0
  • lesson_learned: 0
  • dci_job: 114 (auto-captured)
  • jira_ticket: 156 (auto-captured)

Project Context (2):
  • sami_antilla_manager: Sami Antilla is Fred's manager...
  • franck_baudin_role: Franck Baudin is the Product Manager for DCI...

Tips

  1. Be explicit when stating preferences:
  2. "I prefer pytest over unittest"

  3. "I switched to neovim" (agent will expire old preference and save the new one)

  4. Ask the agent to recall:

  5. "What testing framework do I prefer?"

  6. "What have you learned about DCI?"

  7. Query by time:

  8. "What did you know yesterday at 3pm?"
  9. "What have you learned since Monday?"

  10. "What was happening on June 15?"

  11. Store future plans:

  12. "I'll be at a conference in Berlin on July 10" (stored with future valid_from)

  13. "What are my upcoming plans?" (agent searches with include_future=true)

  14. Correct the agent:

  15. "That lesson about Friday failures was wrong" (agent will retract it)

  16. "I no longer use vim" (agent will expire the preference)

  17. Check periodically what's stored:

  18. "Show me all your stored knowledge"

  19. Or run: python examples/inspect_knowledge_base.py

  20. Backup your knowledge base: 

    cp ~/.ai-assist/knowledge_graph.db ~/backup-$(date +%Y%m%d).db

Learn more

Quick commands

# Interactive mode
ai-assist /interactive

# Ask what agent knows
You: What preferences do you know?
You: What's in your knowledge graph?

# Inspect database
python examples/inspect_knowledge_base.py

# Direct SQL
sqlite3 ~/.ai-assist/knowledge_graph.db "SELECT entity_type, COUNT(*) FROM entities WHERE tx_to IS NULL GROUP BY entity_type;"

That's it! The agent learns automatically. Just use it normally and ask what it knows when curious.