hermes-upgrade-mining-notes

4 min read

Hermes Upgrade — NotebookLM Mining Notes

Notebook: Hermes Agentic OS (31 sources)
URL: https://notebooklm.google.com/notebook/ac4e9066-0267-4317-a772-c9470787d99b
Mined: 2026-07-13

Query 1: Config.yaml Optimizations

Asked: What are ALL the specific config.yaml optimizations mentioned across sources?

Answer:

  • max_bytes: 50000 → 100000+ (prevent truncation of long outputs)
  • compression_threshold: 0.5 → 0.75 or 0.8 (use more context before compressing)
  • target_ratio: 0.2 → 0.4-0.8 (keep more conversation uncompressed after compression)
  • max_line_length: 2000 → 5000 (read long markdown lines)
  • Auxiliary models: Route background tasks to gemini-2.0-flash or deepseek-v4-flash (3-5x cost savings)
  • max_turns: 150 → 60 (prevent token runaway)
  • hard_stop: false → true (auto-terminate on repetitive loops)
  • max_concurrent_children: 3 → 5 (more parallel sub-agents)
  • max_spawn_depth: 1 → 2 or 3 (nested sub-agents)
  • tools.search: auto (lazy-load tool definitions)
  • memory.write_approval: true (stage memory updates for review)
  • skills.write_approval: true (stage skill writes for review)

Query 2: Cron Automation Patterns

Asked: What are ALL the cron job examples and automation patterns?

Answer:

  1. Dreaming Loop — 0 3 * * * nightly, distills sessions into ~/.hermes/holme-profil.md, --deliver local
  2. GitHub Backup — 0 0 * * * midnight, push config/state to GitHub
  3. Morning Interview — reverse-prompt user for priorities, push tasks to Kanban
  4. Kanban Librarian — 0 9 * * 1 weekly, flag tasks stuck >7 days
  5. Heartbeat/Health Check — periodic server monitoring

Query 3: Skills System Techniques

Asked: What are ALL the skills system techniques discussed?

Answer:

  • Skills are plain markdown (SKILL.md) in ~/.hermes/skills/
  • /learn shortcut auto-authors skills from URLs/text
  • Skill stacking: up to 5 skills in one /skill1 /skill2 /skill3 [instruction] command
  • Skill bundles: YAML files in ~/.hermes/skill-bundles/ grouping skills under one slash command
  • Curator lifecycle: prunes, merges, updates skills automatically
  • Self-improvement triggers: ≥5 tool calls, error recovery, human correction, novel discovery
  • Skill taps: publish to GitHub repo, others install via hermes skills tap add
  • Progressive disclosure: Level 0 (index) → Level 1 (full skill) → Level 2 (reference files)

Query 4: Memory & Knowledge Engineering

Asked: What are ALL the memory and knowledge engineering patterns?

Answer:

  • MEMORY.md (2,200 chars) + USER.md (1,375 chars) injected at session start
  • At 80% capacity, agent consolidates with replace tool
  • Dreaming Loop distills nightly into structured profile file
  • Obsidian vault integration turns “dead files” into “living files”
  • Power users symlink Dreaming output to Obsidian for memory graph visualization
  • LLM Wiki pattern for bidirectional knowledge base
  • Context window optimization via compression tuning

Query 5: Multi-Agent Orchestration

Asked: What are ALL the multi-agent orchestration patterns?

Answer:

  • Kanban board: hermes dashboard → Kanban tab, tasks with status/handoffs/workers
  • Profiles as “Pantheon”: Mercury (cheap/cron), Labyrinth (deep research), Philosopher (ambiguous problems)
  • Task decomposition: parent → child via max_concurrent_children and max_spawn_depth
  • Dispatcher pattern: high-reasoning model orchestrates cheaper workers
  • /goal command: verifiable end state, works until outcome achieved
  • Adversarial Loop: Builder (Claude) ↔ Verifier (GPT), loop until no issues
  • Failover coordination: secondary agent monitors/fixes primary
  • Tailscale for remote fleet management across devices
  • Claude OS Bridge: shared memory between Hermes and Claude Code

Query 6: MCP Integration & Hidden Features

Asked: What are ALL the MCP server integration patterns and hidden features?

Answer:

  • High-value MCP servers: Apideck (200+ SaaS), Composio (1000+ apps), AgentQL (DOM extraction), Apollo (B2B leads)
  • supports_parallel_tool_calls: true for read-only servers
  • idle_timeout_seconds: 300 for memory-heavy stdio servers (RAM recovery)
  • tools.search: auto prevents system prompt bloat
  • Hidden features: /learn, /goal, /journey, skill bundles, curator auto-run
  • Security: named API keys, least privilege, separate accounts, gradual access buildup

Applied Config Changes (9 total)

SettingOldNewStatus
compression.threshold0.50.75✅ Applied
compression.target_ratio0.20.4✅ Applied
tool_output.max_bytes50000100000✅ Applied
tool_output.max_line_length20005000✅ Applied
agent.max_turns15060✅ Applied
tool_loop_guardrails.hard_stop_enabledfalsetrue✅ Applied
delegation.max_concurrent_children35✅ Applied
delegation.max_spawn_depth12✅ Applied
skills.guard_agent_createdfalsetrue✅ Applied

Patterns Requiring User Decision (7 total)

  1. Dreaming Loop — nightly cron at 3 AM for session distillation
  2. Morning Interview — reverse-prompting pattern for daily task discovery
  3. Kanban Librarian — weekly triage of stuck tasks
  4. Pantheon Profiles — Mercury (cheap/background) + Labyrinth (deep research)
  5. Adversarial Loop — Builder/Verifier with different models
  6. Skill Bundles — group marketing skills under single slash command
  7. Auxiliary model routing — route background tasks to cheaper model (3-5x savings)

Graph View