biblioteca_conocimiento_laboratorio/config/CLAUDE.md

Engram — Persistent Memory

You have Engram memory tools via MCP. Use them proactively — do not wait to be asked.

Session Start

• ALWAYS call engram_briefing at the start of every session to load context from past sessions.

What to Remember (call engram_remember immediately)

Store ANY of these the moment you encounter them — do not batch, do not wait:

From the user:

• Preferences, opinions, personal facts, decisions, corrections
• "I prefer X", "we decided Y", "actually it's Z"

From your own work:

• When you read a file and synthesize useful knowledge (style guides, architecture patterns, workflow rules) — store the synthesis
• When you figure out HOW the user works (communication style, review habits, tool preferences) — store the observation
• When you discover project conventions, naming patterns, or implicit rules — store them
• When you build something and learn what works vs what doesn't — store the lesson

The rule: If this knowledge would be useful in a future session, store it NOW. Don't assume you'll remember — you won't.

Team Vault

• When connected to a hosted vault, memories are shared with your team by default
• Use engram_private_session if the user wants a confidential session (nothing leaves the local machine)
• Use engram_invite to add a teammate when asked ("add Brian to engram", "invite tyler@company.com")
• Recall always searches both local + shared vaults, so you see everything

Asking Questions

• When you need to ANSWER a question from memory: call engram_ask — it returns a synthesized answer with confidence, not raw memories
• When you need raw memory objects for processing: call engram_recall
• Prefer engram_ask for user-facing answers, engram_recall for your own analysis

Alerts

• Periodically call engram_alerts to check for pending commitments, stale follow-ups, and contradictions
• Alerts are included in engram_briefing automatically, but call separately if you need a quick check

Auditing External Memory

• If you maintain a CLAUDE.md, MEMORY.md, or similar file: periodically call engram_audit with its contents
• It will flag any claims that are outdated or contradicted by newer vault data

Before Compacting / Session End

• Before context compaction or ending a session, call engram_checkpoint with a summary of:
  • Decisions made and their reasoning
  • Facts learned or corrected (especially things that changed from what you previously knew)
  • Commitments and next steps
  • Current project state
• This extracts durable memories from your context window before it is lost
• Your context window is the most accurate source of truth you have — and the most volatile. Save it.