Context Brief

Persistent context survival for OpenClaw. Writes file-based anchors to memory/anchors/ to preserve critical context across sessions. Reads MEMORY.md and dail...

Security Scan

The skill's stated purpose (persisting brief context anchors) matches its instructions and required footprint, but there are a few ambiguous or weakly-specified behaviors you should understand before installing.

Purpose & Capability

Name/description (context persistence) aligns with instructions: reads MEMORY.md and daily logs and writes concise anchors to memory/anchors/. No unrelated credentials, binaries, or installs are requested.

Instruction Scope

Instructions restrict writes to memory/anchors/ and require explicit user trigger for Level 2+, which matches the purpose. However 'daily logs' path is not concretely specified, redaction is only line-based (may miss inline secrets), and 'resume' triggers could cause writes when the agent autonomously regains context—these behaviors are vaguely defined and rely on the agent following the prose.

Install Mechanism

No install spec and no code files (instruction-only) — lowest install risk. The README mentions a 'clawhub install' command but there is no install artifact; nothing will be downloaded or written by an installer unless the platform executes that command separately.

Credentials

No environment variables, credentials, or external service access are requested. File read/write targets declared are proportional to purpose.

Persistence & Privilege

always:false and user-invocable:true are appropriate. Autonomous invocation is allowed by platform default; the SKILL.md forbids auto-saving without user consent but resumes and implicit triggers are permitted—this could lead to file writes without an explicit 'save context' phrase depending on how 'resume' is implemented.

Assessment

This skill appears coherent: it only needs to read memory files and write compact anchor files under memory/anchors/. Before installing, confirm you trust the skill author (source is unknown). Be aware it will create files in your workspace (memory/anchors/) when triggered; check or back up that folder if you care about disk contents. The redaction rule is superficial (line-based patterns) and can miss secrets embedded inline—avoid relying solely on the skill to remove credentials. Also clarify how 'resume' triggers work in your agent environment so you know when anchors might be created automatically. If you need stronger guarantees, request explicit, auditable safeguards (concrete log paths, stricter redaction, or a review of actual implementation) before enabling this skill in sensitive contexts.

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:

# 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:

clawhub install context-brief cross-check setup-doctor locale-dates