Hermes yesterday, October 11, 2026: models stop forgetting their own reasoning

Models stop forgetting their own reasoning. Claude on Bedrock, Gemini, Copilot, OpenAI and more now get their earlier thinking back through every turn, resume and compaction, wherever the provider can read it. 166 commits hit main, here are the 10 that matter, starting with that one.

1. Models keep their own reasoning across turns, resumes and compaction

Before: Hermes sent earlier reasoning back only to a few providers (DeepSeek, Kimi, MiMo, Ollama, OpenRouter), dropped Claude’s signed thinking on Bedrock, Vertex and Foundry, never kept Gemini and Copilot signatures, and compaction deleted reasoning from older turns. After: replay is the default on every route that reads it, through resume, /branch, /steer and compaction, while strict APIs like Mistral, Groq and Cerebras get none. OpenAI’s gpt-5.4 and newer families keep earlier turns’ reasoning, Claude Haiku 5.5 now thinks adaptively, and a provider that rejects a field is retried once without it and remembered for that session. Trade-off: replayed reasoning counts as input tokens. Bedrock, Vertex, Foundry, MiniMax and Copilot were checked against vendor docs, not in live runs.

Pull requests: every provider that can read it, signed reasoning after resume.

2. A subagent’s background job no longer hijacks your chat

Before: when a background process started by a delegate_task subagent finished, the gateway moved the chat onto the subagent’s transcript, ended your real conversation and dropped the parent’s results. It was reported on Discord, Telegram, Slack, Zulip and Feishu and reproduced on all 18 chat platforms. After: the result reaches the conversation that owns the subagent and the chat stays put; a result whose owner cannot be confirmed is retried instead of lost. Subagent process notices stay off by default. A guard for a rarer case, where the subagent’s link to its parent is gone, is still in review.

Pull requests: fix, still in review.

3. Desktop regenerate asks before throwing away later turns

Before: in a long-lived Desktop window, regenerate could aim at a turn hours back and silently archive everything after it; one report lost 107 live messages. After: a regenerate or edit that would archive later user turns asks first, and declining sends nothing. The server only logs unconfirmed deep cuts in this release, so older clients keep working, and refuses them in the next one. The client-side root cause was not reproduced, and a fix for a newer Desktop talking to an older backend is still in review.

Pull requests: fix, still in review.

4. Heads-up: Holographic memory left core early

Correction to yesterday: Holographic was announced to leave core on October 15, but it left core on main yesterday, now that it has a community maintainer (itpartypattaya). Existing users are moved to the catalog plugin by hermes update or on the next agent start, with the same provider name, memory_store.db and settings. With security.allow_lazy_installs: false, agent start installs nothing and prints the hermes plugins install holographic command instead. byterover and retaindb stay bundled until the announced date.

Pull request: Holographic leaves core.

5. persistent_shell: false finally does something

Before: since the spawn-per-call rewrite, terminal.persistent_shell, TERMINAL_SSH_PERSISTENT and TERMINAL_LOCAL_PERSISTENT were read and then ignored: on local and SSH, exported variables always carried over to the next command. After: setting them to false gives each command a fresh login environment, while the working directory still carries over. Defaults are unchanged, so only people who set false see a difference, and Docker, Modal and the other backends are unaffected. The docs now say exactly what carries over: the working directory and exports, not plain variables, functions or aliases.

Pull request: persistent shell setting.

6. browser_exec reaches your LAN without sending it to the cloud

Before: with a cloud browser provider set, browser_exec refused every localhost or LAN URL, and in Browser Use mode it is the only browser tool. When the cloud provider failed (a 429 rate limit, for example), the whole call failed too. After: private URLs run on a local Chromium sidecar, so they never reach the cloud browser; a call mixing public and private URLs runs wholly local; a provider failure falls back to local Chromium. Cloud metadata endpoints stay blocked, and auto_local_for_private_urls: false still blocks private URLs.

Pull requests: LAN via local sidecar, fallback when the cloud fails.

7. Several accounts per connector, each with a name

Before: Hermes assumed one account per hosted app, so a second Gmail could not be told apart, and reconnect could report connected while the broken account stayed broken. After: you can keep home and work accounts, the model picks one per call and every result says which account ran; reconnect repairs one exact account. Manage them in Desktop (Capabilities, Connectors), in the TUI with /connectors and in the CLI with hermes connectors. It relies on the Nous tool gateway side, which the PR asked to ship first, and rename needs a Nous Portal update; until that is live, Hermes says the Portal cannot rename yet.

Pull requests: agent core, TUI and CLI, Desktop.

8. hermes update on main stops tripping over a 403

Before: every hermes update on the main channel first asked a release server for a record that was never published. A 404 fell through to git, but some regional firewalls answered 403 and stopped the update outright. After: main goes straight to the git branch with no extra request. Preview channels still read their records.

Pull request: main channel without 403.

9. Qwen’s content filter ends the turn instead of a 4-minute loop

Before: when Alibaba’s guardrail refused a Qwen request through OpenRouter or the Nous Portal, Hermes treated it as a server error and resent the same content; in a test, one message took 17 requests and 243 s. After: it counts as a content-policy block: 3 requests, 16 s, a fallback model if you configured one, and a message saying the provider’s safety filter refused.

Pull request: Qwen content filter.

10. 2 new plugins in the catalog

Desktop is the day’s theme: both new plugins are for it, and Desktop plugins can now run built-in app actions, like toggling the browser, through ctx.runAction instead of faking a keyboard shortcut (pull request). Only view and navigation actions are allowed. Catalog entries do not wait for a release: they become installable once the published catalog refreshes.

• Desktop: accent-picker, browser-toggle

Meanwhile, in the engine room

Dozens of fixes landed yesterday that you will probably never see. Most of them went into the new connector accounts and the release pipeline, so you never have to meet those bugs.

• A release download that arrives corrupted is fetched again instead of failing the whole run (pull request).
• The plugin scanner stopped flagging 38 catalog plugins on false positives such as README prose, while real findings still flag (pull request).
• Pull requests that add a config key or slash command now need docs, and reviewers see every change to the system prompt and tool schemas (pull request, pull request).
• Stable releases stop building macOS Intel and Termux packages by default; the Termux APT channel stays on the last release that built it (pull request).

If you noticed none of this, it worked.

All on main, ships in the next release. Latest stable: v0.21.7.

Watch the recap

The 26-second motion recap, made with Remotion, is on X and LinkedIn.