Add documents table, session-init hook, and updated memory-extract hook

[auriwren]

Why

AI agents using nova-memory currently face three gaps that this PR closes:

1. File-based memory is invisible across sessions. If an agent creates memory/research-humidifiers.md during one session, future sessions have no way to know it exists unless a keyword search happens to match. There is no registry, no index, no discoverability layer. The documents table solves this: every workspace file gets registered with a path, title, type, and description. An agent can query "what reference docs do I have about the Forge?" instead of guessing filenames.

2. Session context depends on manual file reads. Today, agents load recent context by reading memory/YYYY-MM-DD.md daily log files at session start. This is fragile (files can be stale, missing, or bloated) and puts context hydration outside the database where the structured memories actually live. The session-init hook replaces this by querying PostgreSQL directly for recent events, decisions, and lessons, then injecting them as a bootstrap file. Context comes from the source of truth, not a copy of it.

3. Memory extraction requires manual discipline. Agents are supposed to call memory-db after meaningful conversations to store facts and events. In practice, this gets forgotten. The memory-extract hook automates it: after each assistant response, a heuristic gate checks whether the conversation contains extractable content (new facts about people, events, decisions, lessons). Only when the heuristics match does it call Claude Haiku for structured extraction. A 5-minute cooldown per session prevents extraction storms. The result: memories flow into PostgreSQL automatically without the agent remembering to do it.

Together, these three changes move nova-memory closer to a fully automated memory pipeline where agents write to and read from PostgreSQL as the single source of truth, with files serving as human-readable reference material rather than the memory system itself.

---

What Changed

1. Documents Table (schema.sql)

Added a documents table for tracking workspace files and knowledge documents across sessions. This enables document discovery without filesystem scanning.

• path (unique): Relative file path
• title: Human-readable name
• doc_type: Category (config, memory, tool, hook, skill, etc.)
• description: What the document contains
• tags: GIN-indexed array for fast tag search

2. Session Init Hook (hooks/session-init/)

New hook that fires on agent:bootstrap to inject recent activity context from PostgreSQL:

• Queries events (last 48h), decisions (7d), and lessons (7d)
• Formats as SESSION_CONTEXT.md bootstrap file
• Falls back silently if PostgreSQL is unavailable
• Skips isolated/spawn sessions

3. Memory Extract Hook (hooks/memory-extract/)

Rewritten for OpenClaw compatibility with LLM-based extraction:

• Heuristic gate: Regex patterns detect extractable content before calling LLM
• Cooldown: 5-minute minimum between extractions per session
• Extraction: Claude Haiku parses facts, events, decisions, and lessons
• Storage: Uses memory-db CLI for structured storage
• Safety: Never blocks message delivery; all errors caught and logged

Testing

All three components are running in production on the Auri workspace. Code review findings (command injection fix, schema query corrections, path portability) have been addressed in follow-up commits.