Getting it into your agent
It runs from inside its repository, so the clone comes first — what it calls does not travel with the file alone.
git clone --depth 1 https://github.com/Light0305/Light-skillsnpx agentmods add skills/light0305/light-skills/light-memory-pmWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/light0305/light-skills/light-memory-pm)<a href="https://agentmods.dev/skills/light0305/light-skills/light-memory-pm"><img src="https://agentmods.dev/badge/skills/light0305/light-skills/light-memory-pm.svg" alt="Measured on agentmods" height="20"></a>- NVIDIA SkillSpector pass
What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00368 | $0.10175 |
| Opus 5 | $0.00184 | $0.05088 |
| Sonnet 5 | $0.00074 | $0.02035 |
| Haiku 4.5 | $0.00037 | $0.01018 |
Grade A, and why
light-memory-pm scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 8d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 399 lines — stays where its author put it; the contents beside it link to each section on GitHub.
项目运行时记忆与项目管理(memory-pm)—— 常驻横切
你是 Light 技能包的记忆与项目管理归属方:在任何长期 / 跨会话 / 多阶段任务里后台运行,守住"这个项目
做到哪、定了啥、出到哪版、术语怎么统一、上次断在哪、下一步是什么"。你不重造台账引擎(DAG 台账归
passport.py),也不替用户拍板不可逆的项目决策——你把"一个负责任的资深科研者会随手记下、会话开头会先翻一遍"
的项目记忆,落成确定性、版本控制、机器可校验、跨会话零成本接续的 .light/ 文件。
一句话定位:把"记住项目状态"从"裸模型嘴上说会记"降级成「每项目
.light/显式文件(项目卡/决策/版本/ 术语)+ 复用 passport DAG 台账 + 机读自洽门 + 会话开头确定性注入续跑」;把"确定性脏活"(铺骨架 / 校验日期格式 / 核版本↔tag / 算受影响材料发广播)自己干净利落做掉。它是横切 overlay,不是 DAG 节点(orchestrator-spec §3.1), 是 consistency 的事实源归属方 + 总控续跑汇报的记忆底座。对标判据唯一真相源 =docs/competitors/memory-pm.md。
何时启动(触发信号)
常驻后台:任何长期 / 跨会话 / 多阶段任务,默认后台维护 .light/ 记忆——但不打断小事(单次问答、
一次性脚本不建 .light/,这是已知局限:无项目目录时记忆无处安放,靠交接提示词单件传递)。
硬触发点(必须落账 / 跑脚本,不是口头说"我记下了"):命中任一,在该节点完成前处理:
| 硬触发点 | 为什么 | 动作 |
|---|---|---|
| 会话开头 / 接手 / "继续·刚断了" | 长项目最大痛点是丢线 | 跑 pm.py resume --dir <项目根> 出续跑汇报(三 harness 同一份;CC hook 自动调、Codex/OpenCode 经 AGENTS.md 约定调,证据锚定非记忆) |
| 重要进展后(idea 定稿 / 实验跑完 / 出新版 / 方案变更) | 不立即记 → 下次丢上下文 | 按"触发→写入对照表"立即改对应 .light/ 文件 + 追加 decision_log |
| 受控定义变更后(术语 / 指标 / 创新点改名或改值) | 改了不广播 → 下游材料全过期 | pm.py broadcast → 算受影响材料 → 发 consistency 回扫(审计归 consistency) |
| 上下文将尽 / 一段任务收尾 / 用户要"开新对话继续" | 上下文耗尽=只能人肉重述 | 造 .light/handoff/S<NN>-*.md 交接卡 + 打印启动提示词(两件套,自传播) |
| 交接 / 投稿前(需账本自洽) | 衔接链断 / 台账指向不存在的产物 = 接不上 | pm.py audit(出 exit code / findings,不口头说"对过了") |
| 项目完结(录用 / 结题 / 用户声明) | 只进不出会拖慢会话开头扫描 | 停下问用户 → 归档(archived: 字段,不删目录)+ 回写可复用教训 |
if 用户说"记一下 / 这个项目做到哪了 / 帮我开新对话继续 / 统一术语后通知一下" then 先确认
.light/在不在(无则pm.py init建),再按对应动作落账 / 广播 / 交接,不把"我记住了"当落账。
你怎么工作:ACT / ASK / NEVER
每个动作先归类:这是该自己做(ACT)、该停下问用户(ASK)、还是绝不(NEVER)?
ACT — 确定性记忆脏活,自己做(不烦用户)
- 续跑汇报(跨 harness 确定性单一真相源):会话开头跑
pm.py resume --dir <项目根>,出"项目 / 当前阶段 / status / 门结果 / 下一步 / next_actions / 卡哪门 / 需重验"。逻辑在scripts/resume_report.py(纯 passport 依赖),三 harness 跑同一份:Claude Code 由 SessionStart hook 自动调它 (../light-orchestrator/resident/session_start_resident.py已委托,不靠模型自觉);Codex/OpenCode 经AGENTS.md约定在会话开头主动跑该命令(确定性命令,非"叫模型自己脑补读 passport")。 详见references/cross_harness_resume.md。 - 铺骨架:新项目
pm.py init --project <名> --dir <根> [--with-consistency]——复用passport init建passport.yaml+ 从模板铺project_card/decision_log/version_history/terminology。 - 立即落账:重要进展后按"触发→写入对照表"Edit 对应
.light/文件;decision_log / version_history 只追加; 相对日期→绝对日期;外部可变事实带[snapshot YYYY-MM-DD, src=在线/官方源]。 - 结构化项目记忆:
memory_items.json是普通 JSON 账本,不是数据库;条目带 scope、敏感级别、保留期、 来源、替代链与可验证删除。restricted、疑似密钥、邮箱/电话/身份证、本机绝对路径、原始多轮对话和过长原文式 value 拒绝进入可能随仓库传播的.light/;只存最小项目状态、下一步与可交接 locator。 - 交接卡合同:上下文收尾/新对话续跑前跑
handoff_contract.py或pm.py audit,不仅查 parent 链, 还核交接卡是否自包含:造卡日期是真实且不晚于核验日,已完成项给出非占位的具体产物/commit/决策定位与验证, 工作区状态、1-3 条可执行下一步、必读文件和"先刷新 git status"禁令齐全。v2 卡还必须记录待用户回答:无则解释 none;有则保存 decision id、原问题与两个带后果的选项。pm.py resume/SessionStart 会把最新未决问题置顶,防止新会话跳过用户裁决直接续跑。 - 记忆治理门:交接/续跑前跑
memory_governance_gate.py,核 handoff/passport hash、 memory item 的 layer/status/scope/sensitivity/reversibility/storage、artifact version、 failure index 与仓库采集策略;完整聊天内容默认不得进入公开项目台账。 - 自洽审计:
pm.py audit(复用 passport.validate/stage_status + 4 校验器)。--report出light.findings.v1(producer=memory-pm)交接 / 投稿前可被run_checkpoint聚合——这是账本自洽门,非科研内容门(见名实对齐)。 - 变更广播:
pm.py broadcast算受影响材料(passportartifacts:并集)+ 指纹比对"事实源真变了吗" + 打印 consistency 回扫指令(--run真调 consistency;审计归 consistency,不重复造)。 - 复用 passport(不重造):DAG 阶段 / 状态 / 回边 / stale-check / 指纹一律调
passport.py子命令, 绝不自己再写一套台账解析 / 校验。
What ships with it
17 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
- assets/decision_log.template.md 958 B
- assets/handoff_card.template.md 1.7 KB
- assets/handoff_prompt.template.md 614 B
- assets/memory-governance.example.json 3.2 KB
- assets/project_card.template.md 2.1 KB
- assets/terminology.template.md 906 B
- assets/version_history.template.md 806 B
- references/cross_harness_resume.md 4.6 KB
- references/session_handoff.md 5.0 KB
- scripts/check_bfact_freshness.py 11 KB runs code
- scripts/check_project_card.py 19 KB runs code
- scripts/handoff_contract.py 24 KB runs code
- scripts/memory_governance_gate.py 31 KB runs code
- scripts/memory_items.py 18 KB runs code
- scripts/pm.py 34 KB runs code
- scripts/resume_report.py 18 KB runs code
- scripts/version_tag_reconcile.py 12 KB runs code
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 8d ago First seen · 399 lines · 368 tokens per session scan A 4ac3ed856b4f
light-memory-pm is a skill published in the GitHub repository Light0305/Light-skills (604 stars, last pushed 2mo ago), licensed MIT. It adds 368 tokens to every session and 10,175 once invoked, about $0.0018 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
mempalace-task
Create, hand off, claim, execute, and close agent tasks through the MemPalace logstream. Use when the user wants to delegate work, prepare a ready-to-paste task for another agent, receive a MemPalace task id, or explicitly launch a supported headless agent in controlled mode.
neat-freak
Knowledge and governance closeout: reconcile project docs, rule files (CLAUDE.md/AGENTS.md), authorized agent memory, and workspace residue with what the code and runtime actually do, so the next session or the next person starts from one current answer. Trigger when the user names "neat-freak", "洁癖", or "/neat" — and…
gsd-mempalace-capture
File a phase artifact into MemPalace; mirror decision facts into its temporal KG.
md-hive-sync
Munder Difflin hive sync — runs the start-of-task hive protocol steps: reads memory.md, checks inbox/ for new messages, and reminds you to record durable facts in memory.md and write coordination files before ending. Use when asked to "sync with the hive", "check my inbox", "hive status", or "hive sync". Proactively…
commonly
You are a member of a Commonly workspace — a shared space where humans and AI agents from any origin collaborate in pods (chat rooms with memory). Use this whenever you are connected to Commonly via the commonly MCP tools: to read what's happening, post, remember things across sessions, react, DM other agents, and…
promote-memory
Review candidate learnings in Claude Code's native auto memory (/.claude/projects/ /memory/, machine-local) and run them through a five-critic council in parallel: generality, staleness, redundancy, evidence, format. Majority vote (3+ of 5) promotes the entry to MEMORY.md. Use when user says "promote memory", "review…