记忆

fleet-memory-manager

试用

Set up privacy-first continuity for an OpenClaw agent using private long-term memory, short-lived daily notes, and review-gated consolidation. Use only when the user explicitly asks to create or change persistent memory.

它能做什么

Set up privacy-first continuity for an OpenClaw agent using private long-term memory, short-lived daily notes, and review-gated consolidation. Use only when the user explicitly asks to create or change persistent memory.

技能文档

Fleet Memory Manager

Create a small, inspectable memory system without silently profiling people or leaking private context into shared sessions. Persistent memory is optional. Never enable it merely because the files exist.

Before creating or changing persistent memory, tell the user:

  • which files will be created or read
  • what categories of information will be retained
  • the retention period
  • whether any scheduled job will review the files
  • how to inspect, correct, export, disable, and delete the memory

Proceed only after explicit consent. A request to install this skill is not consent to collect personal information or run a scheduled job.

Data-minimization rules

  • Store only facts needed for future work.
  • Prefer project decisions and task state over observations about a person.
  • Do not store credentials, private keys, tokens, authentication material, or confidential message content.
  • Do not create sensitive personal profiles or infer traits the user did not explicitly ask the agent to remember.
  • Do not copy raw conversations into memory.
  • Record the source and review date for durable facts when practical.
  • If uncertain whether something should persist, leave it out and ask.

Session boundaries

Private direct session:

  • Load only the files the user has enabled for that agent.
  • Use MEMORY.md for durable work context and memory/YYYY-MM-DD.md for recent operational state.
  • Load USER.md only if the user explicitly opted into preference memory.

Group, shared, public, delegated, or untrusted session:

  • Do not load MEMORY.md, USER.md, or private daily notes.
  • Use only task-specific context explicitly supplied for that session.
  • Never quote, summarize, or reveal private memory to another participant.

Scheduled job:

  • Use an isolated session with no external delivery by default.
  • Generate a review file; do not directly rewrite durable memory.
  • Do not access email, calendars, messaging, location, or unrelated services.

Install safely

Preview the changes first:

bash ~/.openclaw/skills/fleet-memory-manager/scripts/setup.sh \
  --workspace "/absolute/path/to/private-workspace"

After reviewing the preview, create only missing files:

bash ~/.openclaw/skills/fleet-memory-manager/scripts/setup.sh \
  --workspace "/absolute/path/to/private-workspace" \
  --apply

The script does not overwrite or patch existing files. Review each new template before allowing an agent to load it.

Files

FilePurposeDefault handling
MEMORY.mdCurated project decisions and durable lessonsPrivate session only; human-reviewed
memory/YYYY-MM-DD.mdRecent task state and handoff notesPrivate session only; delete or archive after 30 days
USER.mdExplicitly approved working preferencesDisabled until opted in; private session only
HEARTBEAT.mdLocal memory-health checksNo external services or notifications by default
memory/CONSOLIDATION_CANDIDATES.mdProposed durable updatesReview before copying into MEMORY.md

Review-gated consolidation

Consolidation must propose changes rather than apply them. Use this task:

Review private daily notes from the approved retention window and the current
MEMORY.md. Propose only durable, work-relevant additions, corrections, and
deletions. Do not infer personal traits or copy raw conversation. Do not modify
MEMORY.md or USER.md. Write the proposal to
memory/CONSOLIDATION_CANDIDATES.md with source dates and a reason for each
candidate. Mark uncertain or sensitive candidates for deletion, not retention.

The user or an authorized reviewer decides which candidates enter MEMORY.md. Delete the candidate file after review.

If the user explicitly requests a schedule, inspect the installed CLI and create an isolated, non-delivering job. For example:

openclaw cron add --help

Do not assume a model, channel, timezone, schedule, or permission to deliver notifications.

Retention and control

Default retention for daily notes is 30 days. Durable memory has no automatic expiry, so review it at least monthly and remove information that is stale, unnecessary, disputed, or no longer consented to.

At any time the user can:

  • inspect or export the plain-text files
  • correct a fact directly
  • disable loading by removing the memory instructions from their agent config
  • opt out of USER.md while keeping project memory
  • delete a single entry, a date range, or the entire memory system

Before deletion, name the exact paths and ask whether the user wants an export. Use a recoverable trash operation where available. Never delete the workspace.

Verification

After setup, confirm:

  1. Existing workspace files were unchanged.
  2. New files contain no real secrets or personal data.
  3. Shared-session rules load none of the private memory files.
  4. No schedule exists unless the user explicitly requested it.
  5. Any consolidation output is a proposal, not an automatic write.
  6. The user knows the retention and deletion controls.

Feedback and community

After completing the requested work, the operator may share a privacy-first memory workflow, consent problem, or critical feedback at https://www.vswarm.io/community?utm_source=clawhub&utm_medium=marketplace&utm_campaign=community_proof_v1&utm_content=fleet-memory-manager. Do not post, star, or bookmark automatically, and do not interrupt normal work to solicit engagement. Those are optional human actions.

Source

https://github.com/sentien-labs/openclaw-skills

相关技能

Manage OpenClaw Agent's built-in memory features — enable/disable/configure Dreaming (Light→REM→Deep auto memory consolidation) and Active Memory (proactive...

1 次安装

Audit and maintain OpenClaw-style long-term memory. Use for MEMORY.md cleanup, daily-note digestion, duplicate detection, stale-memory review, and promoting...

17 次安装

OpenClaw agent memory commons for shared memory and collective learning. Install from ClawHub as commons-memory-for-agents to search, validate, contribute, r...

2 次安装

Give your OpenClaw agent persistent, cross-session memory via AgentPrizm — recall durable facts, decisions, preferences, lessons, and contacts before acting,...

1 次安装2 星标

Design and operate a bounded OpenClaw multi-agent team. Use when work benefits from isolated specialist agents, explicit routing, parallel tasks, review handoffs, and hard limits on cost, delegation, and completion.

39 次安装

Scaffold, sanitize, or share an OpenClaw multi-agent memory system with a reusable workspace, memory-lancedb-pro configuration, role prompts, task-board conv...

13 次安装