project-long-memory

Project Long Memory Skill

面向 Codex / Claude Code / OpenCode 的项目级长期记忆系统，参考 HOPE / Nested Learning 的多层记忆思想。 当前版本已经从“分层记忆仓库”升级为“运行型记忆系统”：

• events/ 与 dialogue/ 负责清醒态捕捉
• bank/ 负责在线工作态的 typed working bank
• dream/ 负责睡眠态的 delta maintenance plan
• startup/ 负责下一次会话启动时的优先激活上下文
• reports/ 负责显式暴露 memory OS 的 bootstrap graph 和 runtime status

目标不是简单堆日志，而是把项目过程沉淀成“可检索、可压缩、可纠错、可在下次启动时快速唤醒”的长期记忆。

当前设计还明确区分三层真相：

• docs/specs/ 与 docs/adr/ 是 law
• .project-memory/ 是 state / evidence
• workflow + startup + dream 是 recovery system

如需一份把 memory / spec / autodream / harness / context compiler / recovery 全部串起来的总纲，请看：

• docs/overview/project-memory-os.md

如需查看 Sabb-openclaw 的类脑模块经验应如何反向迁入、以及哪些明确不该迁，请看：

• docs/specs/sabb-openclaw-selective-backport-v1/requirements.md
• docs/adr/0008-sabb-openclaw-selective-backport-boundary.md

如需查看 Oh My Codex / OMX 应如何融合、以及为什么只做 selective backport，请看：

• docs/specs/oh-my-codex-selective-backport-v1/requirements.md
• docs/adr/0009-oh-my-codex-selective-backport-boundary.md

如需查看把这套系统升级到“一线平台级”应如何推进，请看：

• docs/specs/platform-grade-upgrade-v1/requirements.md
• docs/adr/0007-platform-grade-upgrade-roadmap.md
• docs/architecture/runtime-core-and-storage.md
• docs/ops/compatibility-matrix.md
• docs/ops/operator-handbook.md
• docs/ops/risk-register-and-disable-policy.md
• docs/ops/troubleshooting.md

如需查看 Google / Addy Osmani agent-skills 对本项目的启发、以及为什么只做 adapt 不做全局批量安装，请看：

• docs/overview/2026-04-26-google-agent-skills-lessons-for-project-long-memory.md
• docs/specs/google-skill-github-research-v1/requirements.md

如需查看 Codex Memories、Chronicle、external skills、skill promotion 和 runtime status 的统一风险登记、禁用/回滚策略、attention 降噪规则，请看：

• docs/ops/risk-register-and-disable-policy.md
• docs/specs/risk-register-and-disable-policy-v1/requirements.md

适用场景

• 长期迭代项目，需要跨会话保持上下文
• 需要追溯关键决策、风险、变更和偏好
• 希望记忆可控、可检索、可纠错
• 希望后台整理只生成计划，不直接乱改 durable memory

主要能力

• 四层闭环：capture layer + typed working bank + dream bus + startup pack
• 多层记忆：events/dialogue → summaries/facts/decisions → bank → startup
• 对话记忆：JSONL 对话日志 + 按月全文记录
• 偏好抽取：从对话中提取偏好/规避项
• 话题演进：支持 summaries/<topic>.md
• 反馈闭环：检索反馈与冲突记录
• 再巩固：基于检索频次重写入摘要
• Working bank：world、open_loops、entities、procedures、project_model
• Spec/Task Binding：把 active work 显式绑定到 docs/specs/<feature>/tasks.md
• Startup Pack：每次会话启动时优先读取的 handoff contract
• Dream Bus：replace / delete / compress / startupPriority 计划输出
• Local Dream Backend：可选接入 local-llm-hub 的本地 OpenAI-compatible API
• 多 Agent 标记：自动写入 agent 字段和 tag
• 多角色治理：需求对齐、独立验证、记忆沉淀
• Codex 工作流：会话入口、角色 gate、回忆评测

目录结构

.project-memory/
  events/
    YYYY-MM.jsonl
    archive/
    candidates/
  dialogue/
    YYYY-MM.jsonl
    full/
      YYYY-MM.md
  index/
    events.index
    access.log
    dialogue_access.log
    activation_access.log
  bank/
    world.md
    open_loops.md
    project_model.md
    state.json
    entities/
      index.md
      <entity>.md
    procedures/
      index.md
      <procedure>.md
  dream/
    latest.json
    latest.md
    state.json
    history.jsonl
    plans/
      <timestamp>.json
      <timestamp>.md
  skills/
    candidates/
      latest.json
      history.jsonl
    drafts/
      <skill-name>/
        SKILL.md
        manifest.json
        references/
          provenance.md
    published/
      <skill-name>/
        SKILL.md
        manifest.json
        references/
          provenance.md
  promotions/
    history.jsonl
    latest.json
    applied/
      skills.jsonl
  evals/
    recall.jsonl.example
  reports/
    dream_health_latest.json
    dream_health_latest.md
    bootstrap_graph_latest.json
    bootstrap_graph_latest.md
    runtime_os_latest.json
    runtime_os_latest.md
    context_drift_latest.json
    context_drift_latest.md
    metrics_latest.json
    metrics_latest.md
    operator_health_latest.json
    operator_health_latest.md
    summary_pane_latest.json
    summary_pane_latest.md
    benchmark_summary_latest.json
    benchmark_summary_latest.md
    history.jsonl
  platform/
    metadata_snapshot_latest.json
    metadata_snapshot_latest.md
    history.jsonl
    metadata/
      runtime.db
  runtime/
    current_context.json
    history.jsonl
    checkpoints/
      latest.json
      history.jsonl
      <checkpoint-id>.json
  sessions/
    history.jsonl
  startup/
    stable.json
    delta.json
    latest.md
    latest.json
    context_build.json
    history.jsonl
  summaries/
    short.md
    mid.md
    long.md
    <topic>.md
  facts.md
  policy.md
  feedback.md
  conflicts.md
  decisions/
    YYYY-MM-DD-<slug>.md

Schema files now live under: