🌙 feat: Dreaming — Automatic Background Memory Consolidation

Summary

Add a Dreaming system that automatically consolidates short-term conversation memories into long-term memory during quiet hours, inspired by biological sleep cycles and OpenClaw's Dreaming process.

Motivation

Hermes already has memory nudges, session search, and FTS5 indexing — but memory consolidation is manual and ad-hoc. The agent has no built-in way to:

• Automatically review recent conversations during idle time
• Score and promote important memories to MEMORY.md
• Reflect on patterns and themes across sessions
• Maintain a human-readable "Dream Diary" of insights

This means valuable context from past conversations is often lost unless the user or agent explicitly saves it. A dreaming system closes this gap.

Proposed Design

A new plugins/dreaming/ plugin that runs a 3-phase consolidation cycle via cron:

Scoring system (weighted, 0–1 scale):

• Relevance (30%): meaningful keyword density
• Frequency (24%): how often the topic appears
• Query diversity (15%): appears across multiple sessions
• Recency (15%): newer = higher score
• Consolidation (10%): penalize if already in MEMORY.md
• Conceptual richness (6%): longer, more detailed = richer

Key behaviors:

• Opt-in, disabled by default — users must explicitly enable it
• Respects user activity — skips if user was active within quiet_minutes (default: 60 min)
• Configurable schedule — cron expression, default 0 3 * * * (3 AM daily)
• Dream Diary — human-readable consolidation log in DREAMS.md
• CLI commands — hermes dream run, hermes dream status, hermes dream diary

Config Schema

plugins:
  entries:
    dreaming:
      config:
        enabled: false           # opt-in
        frequency: "0 3 * * *"  # cron schedule
        quiet_minutes: 60        # skip if user active within N minutes
        model: null              # null = default model (for future LLM-based REM)
        max_candidates: 50       # max candidates per cycle
        promotion_threshold: 0.6 # min score to promote
        min_recall_count: 2      # min frequency to promote
        lookback_days: 7         # how many days of sessions to scan
        dream_diary_path: null   # null = ~/.hermes/DREAMS.md

Implementation

I've built a working proof-of-concept in a local branch with:

• plugins/dreaming/__init__.py — main engine (Light/REM/Deep phases)
• plugins/dreaming/cli.py — CLI commands
• plugins/dreaming/tools.py — tool schemas for cron-triggered LLM calls
• plugins/dreaming/plugin.yaml — plugin metadata

Inspiration

• OpenClaw's Dreaming process (PR openclaw/openclaw#19685)
• Human memory consolidation during sleep (hippocampus → neocortex transfer)
• Hermes' existing cron + session store infrastructure

Open Questions

1. Should the REM phase use an LLM call for richer reflections, or is the rule-based approach sufficient for v1?
2. Should dreaming be a built-in plugin or a user-installed one?
3. Any concerns about the scoring weights or thresholds?

Would love to hear if this aligns with the Hermes vision. Happy to iterate on the design and submit a PR if there's interest.