Observations

Observations are LLM-compressed summaries of conversation transcripts. They're the bridge between raw chat history and structured memory — the observe pipeline reads conversations, extracts what matters, assigns priority, and routes findings to the right vault categories.

How Observations Work

Conversation transcript
        ↓
   observe (LLM compression)
        ↓
   Priority assignment (critical / notable / routine)
        ↓
   Auto-routing to categories (decisions/, lessons/, preferences/, etc.)
        ↓
   Stored in observations/ with scored frontmatter

The Pipeline

1. Input: Raw conversation files (OpenClaw sessions, ChatGPT exports, Claude exports)
2. Compression: An LLM reads the conversation and extracts key information — decisions made, lessons learned, preferences stated, commitments given
3. Priority scoring: Each observation gets a priority level
4. Routing: High-value items are auto-routed to typed categories (decisions, lessons, etc.)
5. Storage: The compressed observation lands in observations/ with frontmatter

Scored Importance System

Updated in v2.2.0

Observations use scored importance tags with the format [type|c=confidence|i=importance]:

Available types: decision, preference, fact, commitment, milestone, lesson, relationship, project.

File Structure

Observations are stored in observations/ with date-based filenames:

---
title: "Session 2026-02-15 — Auth system redesign"
priority: 8
source: openclaw/clawdious/main
created: 2026-02-15T14:30:00Z
tags: [auth, security, jwt]
---

## Decisions
- Switching from session tokens to JWT for stateless auth
- Using RS256 signing with key rotation every 90 days

## Lessons
- Session-based auth doesn't scale well with multiple agents
- JWT refresh token flow needs careful error handling

## Action Items
- Implement JWT middleware (assigned to clawdious)
- Update auth docs before v2.5 release

CLI Commands

observe

Run the observation pipeline.

# Watch active OpenClaw sessions
clawvault observe --active

# One-shot compression of a specific file
clawvault observe --compress ~/transcripts/session-2026-02-15.md

# Watch a directory for new files
clawvault observe --watch ~/transcripts/

# Dry run — see candidates without compressing
clawvault observe --active --dry-run

# Run as background daemon
clawvault observe --daemon

reflect

Promote stable observations into weekly reflections — a higher-level summary of patterns across multiple sessions.

clawvault reflect
clawvault reflect --days 7
clawvault reflect --dry-run

Auto-Routing

When the observer compresses a conversation, it doesn't just dump everything into observations/. High-value items are routed to typed categories:

• Decisions → decisions/ (e.g. "We decided to use Postgres instead of MongoDB")
• Lessons → lessons/ (e.g. "Never deploy on Friday without rollback plan")
• Preferences → preferences/ (e.g. "Pedro prefers dark mode in all tools")
• Commitments → tasks or commitments/ (e.g. "I'll have the report ready by Monday")

You can customize routing rules with clawvault route add:

# Route anything mentioning "security" to the decisions category
clawvault route add "security" decisions

# Test a route
clawvault route test "We decided to enable 2FA for all users"