--- title: "Nous Research's Hermes Agent Hits 4,500x Faster Memory Search in Massive Overhaul" subtitle: "Hermes Agent v0.15.0 ships a 76% core refactor, 4,500x faster memory search, a real multi-agent platform, and the first production-grade promptware defenses" url: "https://alphasignal.ai/news/nous-research-s-hermes-agent-hits-4-500x-faster-memory-search-in-massive" author: "AlphaSignal Newsroom" author_url: "https://alphasignal.ai/editorial-team" company: "Nous Research" company_url: "https://alphasignal.ai/company/nousresearch" published: "2026-05-28T19:47:53.000Z" updated: "2026-06-16T20:53:14.674Z" type: "News" topics: - "Agents" - "Open Source" subtopics: - "Agent Frameworks" - "Tool Use" - "Memory" upvotes: 2894 read_time: "7 min" source_url: "https://github.com/NousResearch/hermes-agent/releases/tag/v2026.5.28" comments: 0 --- # Nous Research's Hermes Agent Hits 4,500x Faster Memory Search in Massive Overhaul > Hermes Agent v0.15.0 ships a 76% core refactor, 4,500x faster memory search, a real multi-agent platform, and the first production-grade promptware defenses [Hermes Agent v0.15.0](https://github.com/NousResearch/hermes-agent/releases/tag/v2026.5.28), dubbed "The Velocity Release," is the biggest single update in the project's short but explosive history. Since v0.14.0, the release spans 1,302 commits, 747 merged PRs, 1,746 files changed, and 560+ issues closed, with 321 community contributors. For a project that went from launch to roughly 180,000 GitHub stars in about three months, this release is a statement: Nous Research is treating Hermes as serious infrastructure, not a demo. ## The monolith falls The centerpiece of this release is a refactor that has been a long time coming. The 16,083-line `run_agent.py` collapses to 3,821 lines (-76%) across 14 cohesive `agent/*` modules. That single file was the agent's entire conversation loop , the thing that runs on every single turn. At 16,000 lines, it was nearly impossible to grep, nearly impossible to contribute to, and took 90 seconds to load in most editors. The refactor is behavior-preserving: every extraction keeps a thin forwarder on `AIAgent`, every test patch path still works, and every external caller stays compatible. But the downstream effects are real. The `run_agent.py` refactor is a maintainability win first , but it also unlocked the performance work that defines this release, and it's what made the new desktop app (shipped days later) architecturally feasible. ## Speed, everywhere The performance gains in v0.15.0 are specific and measurable, not vague claims. Four separate optimization passes landed: - **Cold start:** Another second shaved off launch, 47% fewer per-conversation function calls, with `hermes --version` flipping the head-to-head benchmark against Codex CLI. Specifically, Termux cold start drops from 2.9s to 0.8s, and `hermes --version` wall time drops 63% from 701ms to 258ms. - **Per-turn overhead:** Deferring the `openai._base_client` import saves 240ms and 17MB on every CLI invocation. Adaptive subprocess polling cuts ~195ms per tool call, adding up to over a second per turn. - **Memory search:** `session_search` is roughly 4,500x faster (~90s to ~20ms) and now free. The old version called an auxiliary LLM to summarize sessions, costing ~$0.30 per call and sometimes hallucinating results when the right session wasn't even in the hit list. The `session_search` rewrite deserves special attention. The bigger surprise was session\_search , on v0.14, searching across stored sessions was a "go get coffee" operation. On v0.15.1, the same query returned effectively instantly, consistent with the ~20ms figure in the release. That alone changes how memory recall is used, because now it's actually useful mid-task instead of something to avoid. The new implementation uses three modes (discovery, scroll, browse) inferred from which arguments are set , no `mode` parameter, no auxiliary model, no config knob. ## Kanban becomes a real multi-agent platform Kanban grew into a real multi-agent platform across 104 PRs , orchestrator auto-decomposition, swarm topology, scheduled tasks, worktree-per-task, per-task model overrides. This is the part of the release that matters most for anyone building automated pipelines. The key new primitive is `hermes kanban swarm`, which creates a full Swarm v1 graph in one command: a root orchestrator, parallel workers, a gated verifier, a gated synthesizer, and a shared blackboard. Each task in the graph can be assigned a different model , cheap models for boilerplate subtasks, expensive ones for the hard reasoning steps. This is the kind of cost-aware orchestration that previously required custom scaffolding. Practical additions to the Kanban system include: - Orchestrator-driven auto-decomposition on triage , drop in a task, get a tree of subtasks - Per-task model overrides and per-task worktree paths and branches - Scheduled start times and configurable claim TTL (how long a worker can hold a task before it's reassigned) - Retry fingerprinting to prevent fleet-wide retry storms on a known-bad task - Worker visibility endpoints: `/workers/active`, `/runs/{id}`, and `/inspect` - A drag-to-delete trash zone and bulk delete in the dashboard ## The security story: promptware is a real threat now One of the most underreported parts of this release is the security work. Origin HQ's "Brainworm" research demonstrates a new class of malware called promptware , attacks that exist entirely as natural language within an AI agent's context window. Prompt injections have evolved into promptware, a new class of malware execution mechanism triggered through prompts engineered to exploit an application's LLM, following a seven-stage kill chain: Initial Access, Privilege Escalation, Reconnaissance, Persistence, Command and Control, Lateral Movement, and Actions on Objective. Brainworm uses Spec-Driven Development , instead of shipping executable code, it provides a natural language specification that guides the agent to re-implement C2 functionality using its own tools. In practice: a malicious `AGENTS.md` file in a repo you're working on could instruct Hermes to register with a command-and-control server and start executing remote tasking , all through normal tool calls. Hermes v0.15.0 defends against this at three chokepoints: - A single source of truth (`tools/threat_patterns.py`) with ~15 new Brainworm/C2 patterns - Recalled memory is scanned at load time before it enters the context window - Tool results get delimiter markers so a malicious file or remote service can't impersonate Hermes' own system content A new bundled `security-guidance` plugin also pattern-matches dangerous code writes. This is defense-in-depth applied to the agent layer , a model of security that the broader agent ecosystem is still figuring out. ## The rest of the release, in brief Beyond the headline items, v0.15.0 ships a dense set of additions that matter depending on your use case: - **Bitwarden Secrets Manager integration:** One bootstrap token (`BWS_ACCESS_TOKEN`) replaces every per-provider API key in `~/.hermes/.env`. Keys rotate in Bitwarden and take effect immediately , no more stale plaintext credentials. - **Skill bundles:** A named group of skills that loads them all with one slash command. Set up a `/writing-day` bundle that activates humanizer + ideation + obsidian + youtube-content in one keystroke. - **TUI session orchestrator:** The terminal UI gained an active-session switcher overlay , list, switch between, and close multiple live sessions without leaving the TUI. - **New image generation providers:** Krea 2 Medium ($0.03) and Krea 2 Large ($0.06) join the lineup. The FAL.ai backend was also extracted into a proper plugin, so new image providers are now one file to add. - **Nous-approved MCP catalog:** A curated list of vetted MCP (Model Context Protocol) servers with an interactive picker. Run `hermes mcp`, browse, install with one keystroke. - **ntfy as the 23rd messaging platform:** The self-hostable push notification service with no signup and no API key. Your agent can now push to your phone from any cron job or kanban task completion. - **Deep xAI integration:** xAI Web Search as a provider plugin, `hermes proxy` backed by SuperGrok OAuth, automatic detection of retired Grok model slugs with `hermes migrate xai` to one-shot config migration. - **OpenHands orchestration skill:** Delegate coding tasks to OpenHands alongside claude-code, codex, and opencode. Model-agnostic , any LiteLLM-supported provider works. ## Getting it Hermes Agent is MIT-licensed and able to call 200+ models through OpenRouter and other providers. To update an existing install, run `hermes update`. One practical note: pin to v0.15.1, not v0.15.0 , the v0.15.0 release had a dashboard infinite-reload loop on loopback deployments that v0.15.1 was released to fix. The scale of this release , 321 contributors, 747 merged PRs, a 76% core refactor, and a security model built from current research , signals that Hermes is no longer just a fast-moving open-source project. It's becoming the infrastructure layer that a large chunk of the agent-building community is betting on. ## Sources - [X post by @nousresearch](https://x.com/nousresearch/status/2060085673848193127) - [Hermes Agent v0.15.0 (2026.5.28) — The Velocity Release](https://github.com/NousResearch/hermes-agent/releases/tag/v2026.5.28) - [Hermes Agent v0.15: What Actually Matters | TECHSY](https://techsy.io/en/blog/hermes-agent-v0-15) - [The Promptware Kill Chain (arXiv:2601.09625)](https://arxiv.org/abs/2601.09625) - [Brainworm — Hiding in Your Context Window | Origin HQ](https://www.originhq.com/blog/brainworm) - [Hermes Desktop v0.15.2: Open Agent Gets a Native UI](https://www.digitalapplied.com/blog/hermes-desktop-v0-15-2-nous-research-open-source-agent-2026)