context-brief
# Context Brief v3.0
**Install:** `clawhub install context-brief`
Persistent context survival. Captures what matters before it's lost — and restores it when needed.
**Writes ONLY to `memory/anchors/`.** Reads MEMORY.md and daily logs — never modifies them.
## Language
Detect from the user's message language. Default: English.
## How It Works
Three levels of action, based on user request or trigger:
**Level 1 — Context Note** (1-2 unsaved items, low risk)
- Append a brief note to your response, no file writes
- Example: `Unsaved context: [decision X pending, path Y important]`
**Level 2 — Anchor** (3+ unsaved items OR user says "save context")
- Write anchor file to `memory/anchors/`
- Show summary to user after saving
- Anchor contains task state, decisions, pending items — never credentials or secrets
**Level 3 — Deep Anchor** (user explicitly requests "save context deep" or "context check")
- Full anchor with all categories
- User gets the summary
## Triggering
Anchors are written when the **user explicitly asks** ("save context", "context check", "hva husker du", "preserve context") or when resuming after time away.
**Do NOT auto-save without user request.** Level 2+ requires user initiation or explicit consent.
## Write Permissions
| File | Read | Write |
|------|:----:|:-----:|
| `memory/anchors/*.md` | ✅ | ✅ — **Only location this skill writes to** |
| `memory/YYYY-MM-DD.md` | ✅ | ❌ — Never modified |
| `MEMORY.md` | ✅ | ❌ — Recommend changes to user |
| `HEARTBEAT.md` | ✅ | ❌ — Read only |
## Anchor File Format
Write to `memory/anchors/YYYY-MM-DDTHH-mm.md`:
```markdown
# Anchor — YYYY-MM-DDTHH:mm+ZZ:ZZ
## Active State
- Task: [what we're doing]
- Status: [in progress / blocked / awaiting]
## Decisions
- [thing] -> [chosen option] (reason)
## Pending
- [ ] [next action with enough context to execute]
## Key Paths
- [path]: [what it is]
```
**Rules:**
- Create `memory/anchors/` if it doesn't exist
- ISO 8601 timestamps
- Max 20 lines — prioritize ruthlessly
- **REDACT secrets** — strip any line containing: password, token, api_key, secret, bearer, private_key, or credential patterns
- One anchor per checkpoint — never append to old anchors
## After Writing
Show user (in their language):
```
Context anchored (N items) → memory/anchors/YYYY-MM-DDTHH-mm.md
```
## Recovery — On Resume
When resuming or after compaction:
1. Check `memory/anchors/` for recent files (last 48h)
2. If anchor found and items are missing from conversation:
```
Restored from anchor: [brief list of items]
Continuing: [last active task]
```
3. If no anchor and context feels lost, ask the user what they were working on.
## What This Skill Does NOT Do
- Does NOT modify MEMORY.md, HEARTBEAT.md, or daily logs
- Does NOT auto-save without user request or consent
- Does NOT store credentials or secrets
- Does NOT restore silently — always informs user
- Does NOT create files outside `memory/anchors/`
## Guidelines for Agent
1. **Wait for user trigger** — "save context", "context check", or resume
2. **Keep it short** — 20 lines max per anchor
3. **Redact secrets aggressively** — strip any sensitive patterns
4. **Inform on restore** — never restore silently
5. **Only write to `memory/anchors/`** — nowhere else
6. **One anchor per checkpoint** — new file each time
7. **Language follows user** — anchor content in the language being used
8. **Timestamp everything** — ISO 8601
## More by TommoT2
- **cross-check** — Auto-detect and verify assumptions in your responses
- **setup-doctor** — Diagnose and fix OpenClaw setup issues
- **locale-dates** — Format dates/times for any locale
Install the full suite:
```bash
clawhub install context-brief cross-check setup-doctor locale-dates
```
标签
skill
ai