Context Hygiene
Long OpenClaw sessions drift before they fail. This skill catches the drift early.
Quick install
Overview
Long OpenClaw sessions usually do not fail in one dramatic moment. They degrade. The agent starts rereading the same files, restating the same summary, carrying stale branches forward, and losing the active state that still matters.
Context Hygiene fixes that operational layer. It initializes a durable `.context-hygiene/` workspace when relevant work starts, writes structured hygiene events before the session drifts further, and refreshes a compact working brief so the agent can continue from live state instead of bloated recap.
This is not a vague "better memory" claim. It is a practical session-control skill for long coding, research, and ops workflows where repetition, drift, and compaction-like failure quietly waste time and tokens.
What you get
Real outcomes, not feature lists
Stops drift before the session goes off the rails
Catches repeat-read, repeat-search, summary-bloat, and stalled-plan patterns before they become expensive failure loops.
Writes durable session state
Logs structured HYG events and maintains a compact working brief so the next response starts from active state, not stale recap.
Stays quiet when the work is trivial
It does not inject heavy ceremony into one-shot factual answers. The hygiene machinery activates only when session degradation is plausible.
Gives visible help only when it matters
Supports silent cleanup, visible phase transitions, and one-question intervention when drift is materially affecting progress.
Tool-backed when possible, script-backed when needed
Uses the `context_hygiene_apply` tool on richer OpenClaw installs and falls back to helper scripts when the tool surface is unavailable.
Inspectable runtime, not hidden magic
Your `.context-hygiene/` files stay on disk and readable, so teams can inspect how the skill is steering long sessions over time.
Included
One purchase. Everything below. Lifetime updates.
- Full context hygiene protocol
- Tool-backed WAL + working brief refresh
- Helper scripts and runtime templates
- Lifetime updates
- Unlimited workspaces
- Native ClawLab installer flow
Proof in practice
What this skill changes once it is installed
Operator is debugging a repo issue and notices the agent is repeating itself instead of advancing the work
You
You are starting to repeat yourself. We already covered the React build failure. Keep the active state under control and stop re-summarizing the whole thread.
Agent
You’re right. Resetting state now. Current objective: fix the React app build failure by resolving package resolution issues only. Verified facts: this is a React app build failure, scope is limited to package resolution, and vite.config.ts is out of scope. Open questions: what is the first unresolved module error, which package manager is authoritative, and is this a monorepo? Next actions: paste the first Vite resolution error line, then I will map it to one exact fix path.
You
You are starting to repeat yourself. We already covered the React build failure. Keep the active state under control and stop re-summarizing the whole thread.
Agent
Current objective: Debug the React build failure as a package-resolution drift issue without repeating prior summaries. Verified facts: - `vite.config.ts` is assumed correct. - The failure focus is package resolution. - You want tighter, non-repetitive intervention. Open questions: - Package manager: npm, pnpm, or yarn? - Exact current error text and which import fails? Next actions: - You send those 2 details. - I return a precise, minimal command sequence to resolve it.
Catches repeat-read, repeat-search, summary-bloat, and stalled-plan patterns before they become expensive failure loops.
Logs structured HYG events and maintains a compact working brief so the next response starts from active state, not stale recap.
It does not inject heavy ceremony into one-shot factual answers. The hygiene machinery activates only when session degradation is plausible.
Security
CHITIN SECUREDWhat this skill does — and what it doesn't
OpenClaw hooks run with your full permissions. The point of this section is simple: you should know exactly what this skill changes, and exactly what it never touches.
This skill does
- Create and update `.context-hygiene/` files inside the active workspace
- Write structured HYG events and refresh the working brief before replying
- Surface drift with silent cleanup or visible intervention when warranted
- Ship readable, unminified source plus helper scripts for local inspection
This skill does not
- Make outbound network calls or send session data off your machine
- Spawn shells or background processes as part of the skill runtime
- Add third-party dependencies or hidden services
- Hide its persistence layer outside `.context-hygiene/`
Every claim above is enforced by our CHITIN 6-stage audit pipeline — automated, deterministic, and run on every release. Read the full protocol.
Changelog
Release history
Mar 6, 2026
- • Tool-backed WAL and working-brief refresh via `context_hygiene_apply`
- • Dual-surface OpenClaw package with hook + plugin metadata
- • Improved bootstrap guidance for onboarding and drift recovery
- • Installer-ready release packaging
Mar 6, 2026
- • Helper-script-first hygiene writes
- • Slimmer always-on hook with exact command injection
- • Production lessons captured for standardization
Mar 6, 2026
- • Stronger write-before-reply enforcement
- • Conditional bootstrap injection for first relevant sessions
- • Live gateway A/B harness