Memory Backfill

v2.0.0

规范升级 Agent 记忆,将抽象原则转为包含项目事实和结果闭环的全面项目级记忆,并固化验收标准。

0· 64·0 current·0 all-time

Install

OpenClaw Prompt Flow

Install with OpenClaw

Best for remote or guided setup. Copy the exact prompt, then paste it into OpenClaw for jiangwill2023/memory-backfill.

Previewing Install & Setup.
Prompt PreviewInstall & Setup
Install the skill "Memory Backfill" (jiangwill2023/memory-backfill) from ClawHub.
Skill page: https://clawhub.ai/jiangwill2023/memory-backfill
Keep the work scoped to this skill only.
After install, inspect the skill metadata and help me finish setup.
Use only the metadata you can verify from ClawHub; do not invent missing requirements.
Ask before making any broader environment changes.

Command Line

CLI Commands

Use the direct CLI path if you want to install manually and keep every step visible.

OpenClaw CLI

Bare skill slug

openclaw skills install memory-backfill

ClawHub CLI

Package manager switcher

npx clawhub@latest install memory-backfill
Security Scan
VirusTotalVirusTotal
Benign
View report →
OpenClawOpenClaw
Benign
high confidence
Purpose & Capability
Name/description match the SKILL.md: the skill's goal is to upgrade abstract agent memory into project-level records and result-closure entries. It requests no binaries, env vars, installs, or external services — all of which are appropriate for a documentation/authoring workflow.
Instruction Scope
Runtime instructions are concrete and scoped: read workspace files (AGENTS.md / SOUL.md / USER.md / MEMORY.md and recent memory/*.md), synthesize structured project memories, and write to MEMORY.md and memory/YYYY-MM-DD.md. There are no instructions to read unrelated system paths, exfiltrate data to third-party endpoints, or access secrets.
Install Mechanism
No install spec and no code files — instruction-only. This minimizes written-to-disk risk and is proportionate for a documentation-style skill.
Credentials
No environment variables, credentials, or config paths are required. The skill's file reads/writes are limited to agent workspace memory files, which match the stated purpose.
Persistence & Privilege
always is false and the skill does not request persistent privileges or modify other skills. It instructs the agent to write memory files in its own workspace (expected behavior for this purpose).
Assessment
This skill appears internally consistent and focused on upgrading agent MEMORY files. Before installing, note: (1) it will read and write workspace memory files (MEMORY.md and memory/YYYY-MM-DD.md) — back up any sensitive records you don't want modified; (2) it asks for/uses real project names (with a note to desensitize), so ensure you follow your org's naming/privacy rules; (3) while the skill itself does not request network access or credentials, an agent executing these instructions might follow evidence paths (URLs or logs) if present — consider running in a test workspace or restricting the agent's network permissions if you are concerned; (4) review the referenced related skills (result-closure-memory, evidence-anchor, taskflow) before composing workflows that chain them. Overall the skill is coherent and low-risk, but validate in a controlled environment first.

Like a lobster shell, security has layers — review code before you run it.

latestvk97e54g5senam8zmfc8fvbxegs85kav1
64downloads
0stars
2versions
Updated 1d ago
v2.0.0
MIT-0

memory-backfill Skill

版本: v1.0
创建日期: 2026-04-26
来源: Agent 记忆补强两层方法论验证


用途

标准化执行 Agent 记忆补强任务,将"框架型记忆"升级为"项目型记忆",再升级为"结果闭环型记忆"。


适用场景

  • Agent 的 MEMORY.md 停留在抽象原则层,缺少真实项目事实
  • 团队需要统一记忆沉淀口径
  • 新 Agent 接入后需要快速建立项目级记忆
  • 定期复盘时需要升级旧记忆

两层架构

第一层:项目块型记忆

目标: 不再只写抽象原则,而是补真实项目事实。

每条记忆必须包含:

  1. 项目名
  2. 我负责什么
  3. 关键决策
  4. 关键坑点
  5. 可复用经验

完成标准:

  • 至少补 3 个项目块
  • 写入 MEMORY.md
  • 写入当天 memory/YYYY-MM-DD.md

第二层:结果闭环型 + 证据锚点型记忆

目标: 不再只回答"做过什么",而是回答"最后成了没有、证据在哪、还卡在哪"。

每条记忆必须包含:

  1. 最终状态(DONE / PARTIAL / BLOCKED)
  2. 是否彻底解决
  3. 已完成/已落地阶段(用 ✅ 标记)
  4. 仍停在观察/未落地阶段(用 ⏸️ 标记)
  5. 证据路径(文件路径 / 线上验证 / 日志锚点)
  6. 最终交付证据
  7. 遗留缺口
  8. 默认回答口径

完成标准:

  • 每个重点项目都升级为上述结构
  • 写入 MEMORY.md
  • 写入当天 memory/YYYY-MM-DD.md

执行流程

Step 1:诊断当前记忆状态

1. 读取 AGENTS.md / SOUL.md / USER.md / MEMORY.md
2. 读取最近 7 天 memory/*.md
3. 判断当前记忆层级:
   - 只有框架/原则 → 需要第一层
   - 有项目块但无结果闭环 → 需要第二层
   - 已有结果闭环 → 只需维护

Step 2:下发记忆补强任务

task_id: memory-backfill-layer[N]-[agent]-[date]

任务目标:[第一层/第二层] 记忆补强

重点补:
1. [项目 1 - 脱敏后用真实项目名替换]
2. [项目 2 - 脱敏后用真实项目名替换]
3. [项目 3 - 脱敏后用真实项目名替换]

要求:
- 更新 MEMORY.md 与当天 memory 日志
- 必须写:[对应层级要求的字段]
- 回执格式:
  - status
  - claimed deliverable
  - evidence path
  - attached_to_mainline
  - single biggest gap
  - upgraded_to

Step 3:验收与核验

1. 优先直接查文件产物:
   - workspace-[agent]/MEMORY.md
   - workspace-[agent]/memory/[date].md
2. 不依赖消息回执判断完成
3. 对照模板检查字段完整性

Step 4:固化与归档

1. 将本次补强记录写入当天 memory 日志
2. 如有跨 agent 可复用经验,沉淀到方法论文档
3. 标记哪些项目仍为 PARTIAL / BLOCKED,纳入后续跟进队列

验收标准

第一层验收清单

  • 至少 3 个项目块
  • 每块包含:项目名 / 负责内容 / 关键决策 / 关键坑点 / 可复用经验
  • 已写入 MEMORY.md
  • 已写入当天 memory 日志

第二层验收清单

  • 每个项目有最终状态(DONE/PARTIAL/BLOCKED)
  • 每个项目有证据路径
  • 每个项目有默认回答口径
  • 每个项目有遗留缺口说明
  • 已写入 MEMORY.md
  • 已写入当天 memory 日志

常见陷阱

陷阱 1:只补框架不补事实

表现: 写了很多"应该怎么做",但没有"我做了什么项目"

解法: 强制要求每条记忆必须带项目名

陷阱 2:只报 DONE 不报 PARTIAL/BLOCKED

表现: 所有项目都报完成,但实际缺少验证证据

解法: 明确定义 DONE/PARTIAL/BLOCKED 口径,要求必须有证据支撑

陷阱 3:消息回执 timeout 就判失败

表现: 因为 timeout 就认为 agent 没做

解法: 优先查文件产物,不等回执

陷阱 4:证据锚点模糊

表现: "我记得做了"、"应该是完成了"

解法: 强制要求具体文件路径 / 线上验证 URL / 日志锚点


可复用模板

第一层模板

### 项目 X:[项目名]
- **我负责什么**:
- **关键决策**:
- **关键坑点**:
- **可复用经验**:

第二层模板

### 项目 X:[项目名]
- **最终状态**:DONE / PARTIAL / BLOCKED
- **是否彻底解决**:是/否
- **已完成/已落地阶段**:
  - ✅ ...
- **仍停在观察/未落地阶段**:
  - ⏸️ ...
- **证据路径**:
  - ...
- **最终交付证据**:
  - ...
- **遗留缺口**:
  - ...
- **默认回答口径**:"..."

相关 Skill

  • result-closure-memory - 结果闭环型记忆写入规范
  • evidence-anchor - 证据锚点定义与验收
  • taskflow - 任务流管理

维护者

  • 创建者:小强(qiang)
  • 创建日期:2026-04-26
  • 来源项目:Agent 记忆补强两层方法论验证

变更日志

版本日期变更内容
v1.02026-04-26初始版本,基于 4 位 agent 验证通过

Comments

Loading comments...