On April 30, 2026, Nous Research shipped Hermes Agent v0.12.0, codenamed The Curator Release. 1,096 commits, 550 merged PRs, 213 community contributors, and one big idea: your AI agent should clean up after itself. The headline feature is a new autonomous Curator that grades, consolidates, and prunes your skill library on a 7-day cycle. The self-improvement loop got a substantial rewrite. Four new inference providers landed. The 19th messaging platform (Microsoft Teams) shipped. TUI cold start dropped 57%.

If you've been running Hermes for more than a few months, your skill library is probably bloated with near-duplicates, stale drafts, and skills you wrote once and never used again. v0.12 addresses that directly. Skills you don't use get archived. Skills that overlap get merged. Skills you actively touch get prioritized for refinement. The agent maintains its own knowledge base, so you don't have to.

This guide walks through every major v0.12 feature with concrete commands, config snippets, and upgrade steps. For context on what Hermes is and why it differs from session-based agents, see our Hermes Agent developer guide.

1The Curator: Autonomous Skill Library Maintenance

Hermes has always generated skills autonomously. The problem: after six months of heavy use, your skills/ directory looks like a closet you haven't cleaned in years. Duplicates. Drafts. Skills that solved a one-off problem and never got touched again. Skills that drifted from the workflow they were written for.

The Curator is a background agent that fixes this. It runs on the gateway's cron ticker with a 7-day cycle by default, inherits the parent configuration, and has unbounded iterations per cycle. You can pick which model it uses from hermes model under auxiliary.curator, and configure it from the dashboard.

# Check curator status - ranks skills by usage
hermes curator status

# Output includes:
#  - Most-used skills (promoted, active)
#  - Least-used skills (candidates for archival)
#  - Recently consolidated pairs
#  - Per-run reports at logs/curator/run.json and REPORT.md

Each run produces two artifacts you can audit: a machine-readable run.json and a human-readable REPORT.md. Archived skills are classified as either consolidated (merged into another skill) or pruned (removed for lack of use) using a model plus heuristic. Skills you explicitly pin are protected from mutation, and bundled/hub skills have defense-in-depth gates that block the Curator from rewriting them.

Configure the Curator in config.yaml:

auxiliary:
  curator:
    provider: openrouter
    model: deepseek-v4-flash  # cheap aux model is fine
    enabled: true

# Override the default cycle if needed
curator:
  cycle_days: 7
  max_prunes_per_run: 20
  min_skill_age_days: 3  # protect very new skills

2Self-Improvement Loop: The Rubric Rewrite

The Curator handles library-wide maintenance. The self-improvement loop handles per-turn learning: after each conversation turn, a background review fork decides what memories and skills to save or update. v0.12 rewrote this system top to bottom.

What changed:

• Class-first prompt. The old review was a free-form "should we update this skill?" question. The new one is rubric-based with explicit criteria, which stops the model from second-guessing obvious wins.
• Active-update bias. The fork now prefers updating the skill the agent just loaded, and correctly handles references/ and templates/ sub-files inside a skill.
• Runtime inheritance. The fork inherits the parent's live provider, model, and credentials. Previously, config could drift and the review fork would fail silently.
• Scoped toolsets. The review fork is restricted to the memory and skills toolsets. It cannot accidentally touch shell, web, or the filesystem outside its lane.
• Clean shutdown. Memory providers exit properly when the fork finishes, fixing a long-running connection leak on busy gateways.
• Clean context. Prior-turn tool messages are excluded from the review summary so the fork sees the conversation shape, not a wall of JSON.

The practical result: skill updates are faster, more targeted, and less likely to introduce regressions. If you were previously disabling self-improvement because it was noisy or expensive, v0.12 is worth a second look.

3Four New Providers + LM Studio Upgrade

v0.12 adds four new inference providers and upgrades LM Studio to first-class status:

The LM Studio upgrade is the most impactful for local-first users. Previously it rode on the generic custom-endpoint path, which meant no hermes doctor checks, no reasoning transport, and no live /models listing. v0.12 promotes it to a native provider with dedicated authentication, diagnostic checks, proper reasoning content handling, and dynamic model discovery.

Model catalog changes worth noting:

• openai/gpt-5.5 and gpt-5.5-pro now in OpenRouter and Nous Portal catalogs.
• deepseek-v4-pro and deepseek-v4-flash added, with reasoning_content replay preserved across cross-provider histories.
• qwen3.6-plus added to Alibaba-supported models.
• Remote model catalog manifest: OpenRouter and Nous Portal catalogs now pull from a remote manifest so new models appear without requiring a Hermes release.

419 Messaging Platforms: Teams, Yuanbao, Pluggable Gateway

The gateway is now a plugin host. Drop-in messaging adapters live outside the core, and Microsoft Teams is the first plugin-shipped platform. Tencent Yuanbao (元宝) joined natively as the 18th platform, with Teams as the 19th via plugin.

Why this matters: you can now ship a private messaging integration without patching the Hermes core. Build it as a plugin, publish it, and users install it alongside their gateway config. For enterprises running internal messaging platforms (Mattermost forks, in-house chat, domain-specific tools), this is the path forward.

Other gateway changes:

• Native multi-image sending across Telegram, Discord, Slack, Mattermost, Email, and Signal. No more single-image attachments when the model generates a gallery.
• Centralized audio routing with FLAC support and Telegram document fallback for codec mismatches.
• Slack channel_skill_bindings: bind specific skills to specific channels. Your #devops channel only sees infrastructure skills, your #design channel only sees Figma and Claude Design skills.
• Signal native formatting: markdown renders as Signal's bodyRanges, reply quotes work, reactions surface as first-class.
• Plugin hooks: pre_gateway_dispatch, pre_approval_request, and post_approval_response let plugins intercept the approval and dispatch flow.

5Bundled Integrations: Spotify, Google Meet, ComfyUI, TouchDesigner

v0.12 promotes several integrations from optional to bundled-by-default:

Other new bundled or refreshed skills: Humanizer (strips AI-isms from text), claude-design (HTML artifact skill), design-md (Google's DESIGN.md spec), airtable (salvaged with .env API key wiring), pretext (creative browser demos), spike + sketch (throwaway experiments, HTML mockups).

You can now install skills from a direct HTTP(S) URL: hermes skills install <url>. A /reload-skills slash command reloads without a full restart.

# Install a skill directly from a URL
hermes skills install https://example.com/my-skill.zip

# Reload all skills without restarting
/reload-skills

# List enabled and disabled skills
hermes skills list

# Pin a skill to protect it from the Curator and self-improvement
hermes skills pin my-critical-skill

6Performance: 57% Faster Cold Start

Cold start was a persistent complaint in the v0.11 series. v0.12 cut visible TUI cold start by roughly 57% through a set of targeted optimizations:

• Lazy agent init (thanks @OutThisLife) delays the heaviest setup until it's actually needed.
• Lazy imports for OpenAI, Anthropic, Firecrawl, and account_usage. You don't pay the import cost for providers you don't use.
• mtime-cached load_config() avoids redundant config reloads when nothing changed on disk.
• Memoized get_tool_definitions() with TTL-cached check_fn results. Tool discovery gets much cheaper on repeat calls.
• Precompiled DANGEROUS_PATTERNS and HARDLINE_PATTERNS eliminates regex compilation on the hot path.

Other perf work: long-session scrolling is stable (no more TUI jank at 1,000+ messages), virtual history heights are lazily seeded, and Ink text measurements are cached across yoga flex re-passes.

There's also a new configurable prompt_caching.cache_ttl. The 5-minute default stays, but you can opt into a 1-hour TTL for bursty sessions where keeping the cache warm saves real money.

7Security & Reliability Changes

The biggest security-flavored change is also the most contentious: secret redaction is off by default. The old default was causing corruption in patches and API payloads because secret-shaped substrings were being replaced with placeholders, breaking tool outputs that happened to contain those substrings.

Opt back in explicitly:

# config.yaml
redaction:
  enabled: true
  patterns:
    - type: api_key
    - type: aws_key
    - type: ssh_key

Other security and reliability wins:

• Hardline blocklist expanded for unrecoverable commands (rm -rf /, dd to disk devices, destructive git operations).
• [SYSTEM:] marker renamed to [IMPORTANT:] because Azure's content filter was flagging the old marker as a jailbreak attempt.
• Symlinks preserved during atomic file writes instead of being silently replaced with regular files.
• Canonical mask_secret helper used everywhere so logs no longer leak secrets through inconsistent masking.
• Probe /dev/tty by opening it, not by bare existence check. Fixes spurious TTY detection on containerized runners.
• Background reviewer scoped to memory and skills toolsets only. No more accidental web or shell escapes during the review fork.

For deeper security patterns including prompt injection defense, sandboxing, and credential hygiene, see our AI agent security guide and production guardrails playbook.

8New CLI Commands: hermes -z, hermes fallback, update --check

Three new commands are worth adding to your muscle memory:

# Non-interactive one-shot mode (great for CI, scripts, cron)
hermes -z "summarize the git log since last Friday"
hermes -z --model deepseek-v4-flash "run pytest and fix failures"
HERMES_INFERENCE_MODEL=gpt-5.5 hermes -z "review this PR"

# Preview an upgrade before applying
hermes update --check

# Manage provider fallback chain
hermes fallback add anthropic
hermes fallback list
hermes fallback remove openrouter

hermes -z is the killer feature here. It opens up Hermes for CI pipelines, git hooks, and cron jobs without needing a TUI session. Pair it with --model for cheap models on non-critical paths and the full model on release-gate checks.

Additional TUI additions:

• LaTeX rendering in the TUI for math-heavy conversations.
• /reload hot-reloads your .env without restarting.
• Opt-in auto-resume of the most recent session on launch.
• Delete sessions from the /resume picker with d.
• A /mouse toggle that kills ConPTY's phantom mouse injection (fixes the WSL2 ghost-mouse bug).
• A mini help menu on ? inside the input field.

9Migration & Upgrade Walkthrough

The upgrade from v0.11 to v0.12 is a standard hermes update, but a few new safety mechanisms are worth using:

# Step 1: preview the update (new in v0.12)
hermes update --check

# Step 2: opt into the pre-update HERMES_HOME backup
# (add to config.yaml before running update)
update:
  backup_hermes_home: true

# Step 3: run the update
hermes update

# Step 4: if you use redaction for compliance, re-enable it
redaction:
  enabled: true

# Step 5: verify providers still authenticate
hermes doctor

# Step 6: configure the Curator
hermes model
# Pick a cheap aux model for auxiliary.curator

If you're coming from OpenClaw, v0.12 includes several migration hardening fixes:

• Plan-first apply with redaction and pre-migration backup.
• Case-preserving brand rewrite.
• One-time ~/.openclaw residue banner so you know if stale files are still on disk.
• openclaw workspace files resolved from agents.defaults.workspace.
• Model aliases resolved against the real OpenClaw catalog schema.

For a deeper migration walkthrough, see our OpenClaw to Hermes migration guide.

10Breaking Changes & Reverts

A handful of items to be aware of before you upgrade production:

• /provider and /plan slash commands dropped. Use hermes model and hermes fallbackinstead.
• flush_memories removed entirely. Memory flushing now happens automatically on clean shutdown.
• Kanban multi-profile collaboration board reverted. Landed in #16081, rolled back in #16098 pending a design rework.
• computer-use cua-driver reverted. Three preparatory PRs landed then were reverted.
• BOOT.md built-in hook removed. The new hooks tutorial shows how to rebuild the same workflow as a shell hook.
• Secret redaction default flipped to off. Already covered in the security section, but worth repeating.

If you're running Hermes on NixOS, check extraPackages behavior: it now works via per-user profile rather than the system profile, which is a fix but will look different in your package list.

11Why Lushbinary for Your Hermes v0.12 Deployment

We've deployed Hermes Agent in production for healthcare SaaS, e-commerce operations, and developer tooling teams since v0.8. v0.12 is the most meaningful release since the introduction of profiles, and getting the migration right matters for teams running agents at scale.

What we do for teams upgrading to Hermes v0.12:

• Zero-downtime migration: staged upgrades across profiles with rollback paths, compliance-aware redaction re-enablement, and audit logging throughout.
• Curator tuning: dial in cycle cadence, aux model choice, and prune thresholds for your library size. A 50-skill solo setup and a 500-skill team setup need different knobs.
• Provider architecture: pick the right mix of LM Studio (local), GMI Cloud or Azure (hosted), and OpenRouter (fallback) for your latency, cost, and data-residency constraints.
• Plugin development: build the messaging platform, observability tool, or domain skill your org actually needs as a Hermes plugin rather than a core fork.
• AWS production deployment: ECS, RDS for memory, proper IAM boundaries, and infrastructure as code so your agent deployment looks like the rest of your stack.

❓ Frequently Asked Questions

📚 Sources

• Hermes Agent v0.12.0 Official Release Notes (GitHub)
• Curator Feature Documentation (Nous Research)
• Hermes Agent Documentation

Content was rephrased for compliance with licensing restrictions. Release data sourced from the official Nous Research release notes as of May 2026. Features and defaults may change in subsequent releases. Always verify on the official Hermes Agent documentation.