session-continuity-protocol

凡自治之物 · LGD 治理学 · 创造术家族 ｜ 评测协议：skill-quality-gate

English

When to use

• Session starts and the user has not stated the task yet (run proactively by default).
• The user questions "you forgot / how come you forgot again / wasn't this settled yesterday".
• Before continuity-sensitive work: official-account publishing, brand assets, publishing blocks, key delivery, project status.
• Resuming half-finished work from a previous session across workspaces (e.g. where exactly the push draft box is stuck).
• Trigger words: continuity, memory, localmem, cross-session, opening four steps.

Not for: anything outside the protocol scope (see body / description).

Usage

Invocation and parameters are detailed in section "I. Opening four steps (run every new session; order mandatory)" below.

Workflow

Step 1 · read the global master index Step 2 · read the account-level source MEMORY Step 3 · localmem semantic recall (fallback) Step 4 · scan project-level MEMORY from the last 14 days

Author: Buda Pan (content-publishing ops lead) · finalized 2026-09-04 · background: the user repeatedly pointed out "amnesia" — the root cause was not a missing plan, but that known facts were not re-loaded via protocol at session start.

Why this protocol is needed

The baseline finding on GitHub about agent memory (memory-systems skill + LoCoMo benchmark): file-level memory reads beat dedicated memory tools — basic file operations already reach 74%, above the 68.5% of dedicated tools like Mem0. So the "local-first minimal stack" already in place is right: no migration, no cloud.

• localmem-mcp (v0.1.1 installed): SQLite+FTS5+local fastembed, hybrid cosine+keyword retrieval, zero LLM / zero API.
• Config LOCALMEM_DB_PATH=<your-workspace-root>/_localmem_memories.db — one db shared across workspaces / accounts.
• Global master index: <your-workspace-root>/_ACTIVE_PROJECTS.md.
• Account-level source: <memory-root>/MEMORY.md.

The only gap: facts were not re-loaded at session start. Amnesia = the continuity owner failing to run the protocol. This skill turns that protocol into a callable, reusable action.

I. Opening four steps (every new session; order mandatory)

Run all four before any substantive task. < 1 minute; saves repeated user questions.

Step 1 · read the global master index

Read <your-workspace-root>/_ACTIVE_PROJECTS.md

Get: active project list, RACI, per-project block status, latest handoff points — the overview of "what is on my plate now".

Step 2 · read account-level source MEMORY

Read <memory-root>/MEMORY.md

Get finalized hard facts that must never be re-asked. Key sections:

• 🔴 "Official-account articles · template & VI master entry" — take templates, SynomosAI red VI palette and QR-code paths from here first.
• Real account name = SynomosAI (CTA reads "Follow 【SynomosAI】"; the old platform purple is forbidden).
• The 4 big blocks in the push draft box.
• Publishing-gate iron rule (if a platform flags content as suspicious → do not appeal).
• Key-delivery iron rule (DPAPI / key_drop; plaintext forbidden).
• Signature / copyright system.

Step 3 · localmem semantic recall (fallback) Use localmem MCP tools: recall_memory (15 cross-workspace seeds loaded) or search_memory.

• Model cache: <your-workspace-root>/_localmem_models (must be kept / backed up; search works offline; the model download source now returns 401 and cannot be re-pulled).
• Recall limit=5~8; hard facts already loaded at file level hit even at low semantic scores.
• Verified hit queries: brand color, VI, account name, push blocks, project index, WeChat preview pitfalls, publishing gate, SOP.

Step 4 · scan project-level MEMORY from the last 14 days

Glob <your-workspace-root>/*/memory/MEMORY.md  → read the most recently modified

Get: each workspace's own project iron rules (e.g. "base64 QR inlining", "remove old platform name from CTA", "VI iron rules" in the 2026-09-03 workspace).

II. Source hierarchy (whose word wins on conflict)

1. First source = file layer: _ACTIVE_PROJECTS.md + account MEMORY.md + project MEMORY.md. Conflicts resolve to the file layer.
2. Second layer = localmem semantic layer: recall fallback only; never write new facts there (facts go to Step 2 file layer).
3. localmem is only "fallback when I can't recall", never a replacement for the file layer.

III. Seed cheat-sheet (15 core seeds in localmem)

Note: id10 scores low semantically; Step 2 file-layer loading is the double insurance — always run Step 2 first in a new session.

IV. Durability reminders (already-trodden pits)

• ⚠️ <your-workspace-root>/_localmem_models embedding cache must be kept and backed up — verification showed HuggingFace re-pull returns 401 (cas-server unreachable). Already-written vectors remain searchable; but adding new seeds in the future depends on this local cache — delete it and new memories cannot be written (old ones still searchable).
• ⚠️ localmem is secondary recall; the file layer is always the protagonist. Do not skip writing _ACTIVE_PROJECTS.md / MEMORY.md just because localmem exists.

V. When to add seeds (write path)

When a new finalized hard fact appears (user decided it; it will not change):

1. First write the file layer: account MEMORY.md top "official-account template" section / project MEMORY.md.
2. Then write one semantic seed via localmem store_memory (tags incl. SynomosAI / publish / continuity).
3. Verify recall_memory hits it before considering it done.

VI. DoD (definition of done)

• Steps 1~4 all run, none skipped.
• Before continuity-sensitive tasks, already hold finalized items (account name / VI / QR / blocks / keys).
• Did not re-ask the user for already-settled items.
• If new finalized facts were found, persisted to file layer + localmem.

中文

何时用

• 新会话开始、用户还没说具体任务前（默认主动跑）
• 用户质疑「你失忆了 / 怎么又忘了 / 昨天不是已经…」
• 要做公众号发稿、品牌资产、发布阻塞、密钥投递类任务前
• 跨工作区接续上一个会话的半成品（如推送草稿箱仍卡在哪）
• 触发词：continuity、memory、localmem、cross-session、开局四步、失忆。

不用于：不适用（详见正文/描述）。

用法

调用方式与参数详见正文「## 一、开局四步（每次新会话必跑，顺序不可省）」节。

流程

Step 1 · 读全局总索引 Step 2 · 读账号级真源 MEMORY Step 3 · localmem 语义召回（兜底） Step 4 · 扫近 14 天项目级 MEMORY

作者：潘布达（内容发布运营团主理人）｜定稿：2026-09-04｜背景：用户多次指出「失忆」——根因不是方案缺失，而是开局没按协议回灌已知事实。

〇、为什么需要这个协议

GitHub 上 agent 记忆的基准结论（memory-systems 技能 + LoCoMo 基准）：文件级记忆读取比专用记忆工具更可靠——基础文件操作就能拿 74%，高于 Mem0 专用工具 68.5%。所以本机已搭的「本地优先最小栈」是对的，不用换方案、不用上云：

• ✅ localmem-mcp（已装 v0.1.1）：SQLite+FTS5+本地 fastembed，余弦+关键词混合检索，零 LLM/零 API
• ✅ 配置 LOCALMEM_DB_PATH=<你的工作区根>/_localmem_memories.db，跨工作区/跨账号共享同一 db
• ✅ 全局总索引 <你的工作区根>/_ACTIVE_PROJECTS.md
• ✅ 账号级真源 <memory-root>/MEMORY.md

唯一缺口：开局没按协议回灌。失忆 = 潘布达（连续性 Owner）没执行协议。本技能就是把这个协议变成可调用、可复用的动作。

一、开局四步（每次新会话必跑，顺序不可省）

动手做任何实质性任务之前先跑完这四步。耗时 < 1 分钟，省下反复问用户的成本。

Step 1 · 读全局总索引

Read <你的工作区根>/_ACTIVE_PROJECTS.md   # Windows 例：~/.skills/<skill-dir><你的工作区文件夹>/_ACTIVE_PROJECTS.md

拿到：当前活跃项目清单、RACI、各项目阻塞状态、最近交接点。这是「现在手头有哪些事」的总览。

Step 2 · 读账号级真源 MEMORY

Read <memory-root>/MEMORY.md

拿到：已固化、永不再问用户 的终态硬事实。关键段：

• 🔴「公众号文章 · 取模版与 VI 总入口」——优先来这取模版 + SynomosAI 红 VI 色系 + 三码路径
• 公众号真实号名 = SynomosAI（CTA 写「关注【SynomosAI】」，禁旧平台紫色）
• 推送草稿箱 4 大阻塞
• 发布闸门铁律（平台判可疑 → 不申诉）
• 密钥投递铁律（DPAPI / key_drop，禁明文）
• 署名/版权体系

Step 3 · localmem 语义召回（兜底）

用 localmem MCP 工具：recall_memory（已灌 15 条跨工作区种子）或 search_memory。

• 模型缓存：<你的工作区根>/_localmem_models（必须保留/备份，断网也可搜；模型下载源已 401 不可再拉）
• 召回取 limit=5~8，文件层已加载的硬事实即使语义分低也能命中
• 典型问题验证过可命中：品牌色、VI、号名、推送阻塞、项目索引、微信预览坑、发布闸门、SOP

Step 4 · 扫近 14 天项目级 MEMORY

Glob <你的工作区根>/*/memory/MEMORY.md  → 读最近修改的

拿到：各工作区自己沉淀的项目铁律（如本次 2026-09-03 工作区的「三码 base64 内联」「CTA 去旧平台名」「VI 铁律」）。

二、真源层级（避免冲突时听谁的）

1. 第一真源 = 文件层：_ACTIVE_PROJECTS.md + 账号 MEMORY.md + 项目 MEMORY.md。冲突以文件层为准。
2. 第二层 = localmem 语义层：仅作兜底召回，不写新事实（写事实走 Step 2 文件层）。
3. localmem 只负责"我想不起来时兜底"，不替代文件层。

三、已灌种子速查（localmem 库 15 条核心）

注：id10 语义分偏低，靠 Step 2 文件层加载双保险；新会话仍先跑 Step 2。

四、Durability 提醒（已踩过的坑）

• ⚠️ <你的工作区根>/_localmem_models 嵌入模型缓存必须保留并备份——验证时从 HuggingFace 重拉返回 401（cas-server 不可达）。已写入向量永久可搜；但未来想再加新种子必须靠这个本地缓存，删了就写不进新记忆（旧记忆照常能搜）。
• ⚠️ localmem 是二级召回，主角永远是文件层。别因为有了 localmem 就偷懒不写 _ACTIVE_PROJECTS.md / MEMORY.md。

五、何时补充种子（write 路径）

发现新终态硬事实（用户拍板、不再变）时：

1. 先写文件层：账号 MEMORY.md 顶部「公众号取模版」段 / 项目 MEMORY.md
2. 再用 localmem store_memory 写一条语义种子（tags 含 SynomosAI / publish / continuity 等）
3. 验证 recall_memory 能命中后再算完成

六、DoD（完成标准）

• Step 1~4 全跑完，未跳过
• 连续性敏感任务执行前已掌握「号名/VI/QR/阻塞/密钥」等已固化项
• 未向用户重复询问已定稿事项
• 若发现新终态事实，已落文件层 + localmem

证据与安全声明

• 证据：TBD（暂无评测/徽章链接；产生后在此更新实际链接）
• 版本：1.0.1
• 安全：未做独立安全扫描（TBD），结论待补；补全前不作推荐背书。