Hermes Agent just shipped a kill switch for autonomous memory writes. Enable it now.

Self-improving agents have a silent failure mode.

They remember things wrong. They write skills that drift. And until today, there was no gate between "the agent learned something" and "it's permanently in your profile."

Hermes Agent shipped Write Gate on June 10.

Two booleans. `memory.write_approval` and `skills.write_approval`. Flip either to `true` and every memory update, skill creation, and skill edit gets staged instead of applied.

Your agent drafts a memory write. It sits in `/memory pending`. You review it. You approve or reject. Same for skills — `/skills pending` shows a one-line gist, `/skills diff <id>` shows the full unified diff.

Why this matters:

The existing approval system only gated dangerous shell commands. Memory writes. Skill creation. Tool description edits. All ran without confirmation.

If your agent was running on a smaller model that hallucinated context, or if you were in a regulated environment where unapproved changes carry risk — you had zero control.

Now you do.

The config is two lines:

```
memory:
write_approval: true
skills:
write_approval: true
```

Foreground memory writes prompt inline in the CLI. Background self-improvement reviews and all skill writes stage for review. Nothing touches your profile until you say yes.

The commit (`095f526`) replaced a confusing tri-state `write_mode` (on|off|approve) with a clean boolean. The old "on" state looked like gating was active when it wasn't. Now `false` means write freely, `true` means ask first.

Run `hermes update` to get it today. It ships in the next major release.

If you're running Hermes Agent in production — or anywhere an autonomous agent touches persistent state — enable this now. The cost of a wrong memory write compounds across every future session.

SOURCE: https://hermes-agent.nousresearch.com/docs/user-guide/features/memory#controlling-memory-writes-write_approval
VERIFIED: NousResearch/hermes-agent commit 095f526, Digg.com post (98K views), Hermes Agent documentation
SIGNAL: Self-improving agents need self-restraint mechanisms. Write Gate is the first production-grade approval layer for agent memory and skill mutation.