In-session write → background consolidation → promotion / retirement → audit — a fully automated knowledge loop piggybacked on bg_review对话内写入 → 后台巩固 → 晋升/淘汰 → 审计——全部搭载 bg_review 的全自动知识闭环
Self-learning is a closed loop — write in-session, consolidate in the background, promote or retire, audit everything. There are no resident processes: every step piggybacks on bg_review, the background review with a 6-hour cooldown, and each step carries its own budget and kill switch.self-learning 是一条闭环：对话内写入、后台巩固、晋升/淘汰、全程审计。无常驻进程——每一步都搭载在 bg_review（6 小时冷却的后台复查）上，且每一步都有独立预算与 kill switch。
┌──────────── 反馈环：MEMORY.md 快照 + topic-index + skills 注入下一轮对话 ◄────────────┐ ▼ │ 对话轮次 —— L1 实时写入 · L2 每轮结论提取 · L3 空闲 flush（预算 ≤2 memory + ≤5 topic） │ ├─► memory ─── 近重复？─┬─ 是 ─► NOOP 拒写 │ │ └─ 否 ─► MEMORY.md 写入（+ ledger）─────────────────────────────┤ ├─► topic save ─┬─► ADD ─────────► history.jsonl（append-only） │ │ ├─► SUPERSEDE ──► 旧条目软失效（invalid_at），修正链 +1 ──┐ │ │ └─► NOOP（近重复，不写） │ │ └─► skill_manage 创建 / patch ──► skills/（created_by 标记） │ │ 链深 ≥2 且存活 ≥3d = 晋升候选 │ │ │ L8 bg_review 总线（6h 冷却，空闲搭载，各环节独立 due-check、fail-silent） │ │ ├─► L5 wiki lint（每 7d）─── 冲突 ──► conflicts.jsonl（只记录，不推送） │ │ ├─► L4 每日蒸馏（per-topic watermark，≤5 topic/轮）──► summary.md 重写 │ │ └─► L6 每周晋升：规则门（免费）──► LLM 门（单次）◄───────────────────────┘ │ │ │ ├─► promote（≤2 写/轮）──► MEMORY.md ───────────────────────────┘ ├─► skip（终局判决，不再烧预算） └─► deferred（写失败 / 超预算）──► 重试环：下轮重新入队 ──► LLM 门 L7 Skill Curator（闲置 ≥90d）──► skills/.archive/ ◄──► 可逆：移回目录 + refresh 即恢复 L9 审计贯穿所有箭头：ledger · supersede 链 · distill / promotions / conflicts.jsonl · telemetry ┌──────────── feedback loop: MEMORY.md snapshot + topic-index + skills injected next turn ◄───────────┐ ▼ │ Conversation turn — L1 realtime write · L2 per-turn extraction · L3 idle flush (≤2 memory + ≤5 topic) │ ├─► memory ─── near-dup? ─┬─ yes ─► NOOP, write rejected │ │ └─ no ──► MEMORY.md write (+ ledger) ────────────────────────────────────────┤ ├─► topic save ─┬─► ADD ─────────► history.jsonl (append-only) │ │ ├─► SUPERSEDE ──► old entry soft-invalidated (invalid_at), chain +1 ──┐ │ │ └─► NOOP (near-duplicate, not written) │ │ └─► skill_manage create / patch ──► skills/ (created_by marker) │ │ chain ≥2 and age ≥3d = promotion candidate │ │ │ L8 bg_review bus (6h cooldown, idle piggyback; per-stage due-check, fail-silent) │ │ ├─► L5 wiki lint (every 7d) ─── conflicts ──► conflicts.jsonl (record only, no push) │ │ ├─► L4 daily distill (per-topic watermark, ≤5 topics/run) ──► summary.md rewrite │ │ └─► L6 weekly promotion: rule gate (free) ──► LLM gate (one call) ◄──────────────────┘ │ │ │ ├─► promote (≤2 writes/run) ──► MEMORY.md ─────────────────────────────────────┘ ├─► skip (terminal verdict, never re-burns budget) └─► deferred (write failure / over budget) ──► retry loop: re-queued next run ──► LLM gate L7 Skill Curator (idle ≥90d) ──► skills/.archive/ ◄──► reversible: move back + refresh to restore L9 audit across every arrow: ledger · supersede chain · distill / promotions / conflicts.jsonl · telemetry
Figure 1图 1 — The real topology: three loops and three forks. Loops — the feedback loop (stores injected back into the next turn), the correction chain (SUPERSEDE deepens a chain until it qualifies for promotion), and the retry loop (deferred promotions re-queue next run). Forks — topic save (ADD / SUPERSEDE / NOOP), the bg_review bus (L5 / L4 / L6), and the LLM gate verdict (promote / skip / deferred).真实拓扑：三个环、三处分叉。环——反馈环（存储注入下一轮对话）、修正链（SUPERSEDE 不断加深链条直至满足晋升条件）、重试环（deferred 的晋升候选下轮重新入队）。分叉——topic save 三判决（ADD / SUPERSEDE / NOOP）、bg_review 总线三环节（L5 / L4 / L6）、LLM 门三判决（promote / skip / deferred）。
Realtime writes the agent makes proactively during a conversation, through three tools plus a routing discipline.会话期间 agent 主动发起的实时写入：三个工具加一条路由纪律。
memory — MEMORY.md (agent notes, 3000-char hard cap) + USER.md (user profile, 2000 chars). Actions: add / replace / remove / reinforce. Near-duplicate detection (Jaccard similarity) returns NOOP and rejects the write; reinforce stamps an entry with a ⟦×N⟧ reinforcement count; every write goes into the audit ledger. Snapshot freeze: a snapshot is loaded at session start and never hot-updated, protecting the prompt cache.memory — MEMORY.md（agent 笔记，3000 字符硬上限）+ USER.md（用户画像，2000 字符）。动作：add / replace / remove / reinforce。近重复检测（Jaccard 相似度）→ NOOP 拒写；reinforce 给条目打 ⟦×N⟧ 强化计数；所有写入进审计 ledger。快照冻结策略：会话开始时载入快照、不热更，以保护 prompt cache。
topic — at most 10 topics; each is a summary.md (≤1000 chars, LLM-distilled) plus a history.jsonl (append-only fact stream). save returns one of four decisions — ADD / SUPERSEDE / NOOP / ERROR. supersede is a soft invalidation: the old entry is stamped with invalid_at plus a superseded_by pointer, never deleted, forming a replayable correction chain. Every fact carries a stable entry_id.topic — 最多 10 个 topic，每个 = summary.md（≤1000 字符，LLM 蒸馏）+ history.jsonl（append-only 事实流）。save 返回 ADD / SUPERSEDE / NOOP / ERROR 四种决策。supersede 是软失效：旧条目打 invalid_at + superseded_by 指针，永不删除，形成可回放的修正链。每条事实都有稳定的 entry_id。
skill_manage — creates and patches skills (procedural knowledge) in-session. Agent-built skills carry an exclusive created_by="skill_manage" marker that the Curator later recognizes.skill_manage — 对话内创建 / patch 技能（程序性知识）。agent 自建技能带 created_by="skill_manage" 独占标记，供后续 Curator 识别。
Six-way routing discipline六路路由纪律 — every fact is routed to exactly one home:每条事实路由到唯一归属：
| Fact事实类型 | Home归属 |
|---|---|
| Name / timezone姓名 / 时区 | user_settings |
| User preferences用户偏好 | USER.md |
| Agent identityagent 身份 | agent_profile |
| Reply style回复风格 | SOUL.md |
| Reusable workflows可复用工作流 | skill |
| Environment facts / lessons环境事实 / 教训 | MEMORY.md |
| Specific facts (address / hash / threshold)具体事实（地址 / 哈希 / 阈值） | topic |
Two lightweight capture layers turn conversation into gated knowledge without touching the main model's budget.两个轻量沉淀层，把对话转化为受门控的知识，不动主模型的预算。
L2 — per-turn extraction. After each turn ends, verified conclusions are extracted at low cost (commit 705c67b2) without involving the main model.L2 — 每轮轻量提取。turn 结束后提取已验证结论（commit 705c67b2），低成本，不动主模型。
L3 — Flush. When the session goes idle, flush fires: an independent small model with tools reviews the recent conversation and routes what is worth keeping into the right stores. It enforces a hard write budget — ≤2 memory adds and ≤5 topic saves per flush — and prefers topics over memory. The flush topic tool is a write-append-only wrapper: it can only save, never read-modify-delete.L3 — Flush。会话空闲时触发 flush：独立小模型带工具审读近期对话，把值得保留的内容路由写入。写入预算强制：每次 flush ≤2 条 memory add + ≤5 条 topic save，且倾向 topic 而非 memory。flush 的 topic 工具是 write-append-only 包装（只能 save，不能读改删）。
Verification ladderVerification ladder — unverified guesses ("maybe X?") never enter storage; only verified facts are written, distilled into reusable rules rather than raw observation logs.未验证的猜测（"maybe X?"）禁止入库；只写已验证事实，且要求蒸馏成可复用规则而非原始观察日志。
Pre-compact flush contractPre-compact flush 契约 — flush is forced before compaction (commit 237ab1e5), so context loss during compression never evaporates knowledge.压缩前强制 flush（commit 237ab1e5），防止上下文丢失导致知识蒸发。
Daily distillation (PR #851, core/memory/distill.py) piggybacks on the bg_review full pass, at most once per day via wall-clock due-check.每日蒸馏（PR #851，core/memory/distill.py）搭载在 bg_review full pass，经 wall-clock due-check 每日最多一次。
Per-topic watermark. Only topics that gained history entries since their own last successful regen are re-distilled; mtime is used for ordering only.Per-topic watermark。只重蒸馏「自己上次成功 regen 后新增了 history 条目」的 topic；mtime 仅用于排序。
Budget. ≤5 topic regens per round (distill_max_topics, configurable); topics that exceeded the budget or failed keep their watermark and are retried next round.预算。每轮 ≤5 个 topic regen（distill_max_topics 可配）；超预算 / 失败的 topic 保留 watermark，下轮重试。
Logs & telemetry. logs/wiki/distill.jsonl; telemetry event memory.distill.日志与遥测。logs/wiki/distill.jsonl；遥测事件 memory.distill。
Weekly wiki hygiene (core/memory/wiki_maintenance.py) — every 7 days, autosuspend-safe: no cron, no resident process, pure piggyback plus a wall-clock due-check.每周 wiki 卫生（core/memory/wiki_maintenance.py）——每 7 天一次，autosuspend-safe：无 cron、无常驻进程，纯搭载 + wall-clock due-check。
The cheap pure-rule gate runs first, LLM spend runs after; conflicts are only written to conflicts.jsonl, never auto-pushed. A 180s collaboration circuit breaker bounds the pass (MAINTENANCE_TIMEOUT_S).便宜的纯规则门在前，LLM 花费在后；发现冲突只写 conflicts.jsonl，不推送。180s 协作熔断（MAINTENANCE_TIMEOUT_S）封顶。
New in PR #1164 (core/memory/promotion.py, 2026-08-12): a weekly promotion pipeline that closes the gap between topic — the high-capacity fact store — and MEMORY.md — the small behavioral store injected every turn — which previously had no machine path.PR #1164 新增（core/memory/promotion.py，2026-08-12）：每周晋升管线，补齐 topic（高容量事实库）→ MEMORY.md（每轮注入的小型行为库）此前只有人工通道的缺口。
Four stages: rule gate (free) → LLM gate (single bounded call) → MemoryStore.add → audit log. The final semantics: a decision is final — it never re-burns budget.管线四段：规则门（免费）→ LLM 门（单次有界调用）→ MemoryStore.add → 审计日志。终局语义：判决终局，不重复烧预算。
① Rule gate (free)① 规则门（免费） — candidates = surviving facts && supersede chain depth ≥2 (corrected at least once, proving ongoing maintenance) && survived ≥3 days && never decided. ≤8 candidates per round.候选 = 存活事实 && supersede 链深 ≥2（至少被修正过一次，证明被持续维护）&& 存活 ≥3 天 && 未决策过。每轮 ≤8 候选。
② LLM gate (single bounded call)② LLM 门（单次有界调用） — gemini-3.6-flash judges candidates one by one against the current MEMORY.md snapshot: promote — rewritten into a generalized rule of ≤220 chars, keeping key numbers / IDs / paths — or skip (topic-local detail / transient / near-duplicate). The verdict is final.gemini-3.6-flash 对照当前 MEMORY.md 快照逐个判决：promote（改写成 ≤220 字符的泛化规则，保留关键数字 / ID / 路径）或 skip（topic 局部细节 / 瞬态 / 近重复）。判决终局。
③ Write (MemoryStore.add)③ 写入（MemoryStore.add） — ≤2 entries per round (aligned with the flush budget); MEMORY.md near-duplicate = final skip; write failure (full / error) = deferred, retried next round; over budget = deferred_budget, retried next round.每轮 ≤2 条（与 flush 预算对齐）；MEMORY.md 近重复 = 终局 skip；写失败（满 / 错误）= deferred，下轮重试；超预算 = deferred_budget，下轮重试。
④ Audit④ 审计 — promotion_done_ids (FIFO, capped at 400) lives in review_state; the audit log is logs/wiki/promotions.jsonl — the equivalent of OpenClaw's Dream Diary; telemetry memory.promotion; kill switch promotion_enabled. Tested with 6 cases (rule gate / exclusion / end-to-end / due gate / write budget / LLM degradation) on a real MemoryStore end-to-end.promotion_done_ids（FIFO 上限 400）存 review_state；审计日志 logs/wiki/promotions.jsonl（相当于 OpenClaw 的 Dream Diary）；遥测 memory.promotion；kill switch：promotion_enabled。测试：6 用例（规则门 / 排除 / 端到端 / due 门 / 写预算 / LLM 降级），真 MemoryStore 端到端。
Skill hygiene (PR #592 + #661, core/skills/curator.py + scripts/skill_curator.py): agent-built skills — one-off helpers, near-duplicates — accumulate and eat system-prompt budget while diluting skill selection.技能卫生（PR #592 + #661，core/skills/curator.py + scripts/skill_curator.py）：agent 自建技能（一次性 helper、近重复）堆积会占系统提示预算、稀释技能选择。
Safety model — all four must hold安全模型——4 条全满足才是候选 — created_by=="skill_manage" (official / preinstalled / marketplace skills are never touchable); not pinned; idle beyond idle_days (default 90 days); the directory really exists.created_by=="skill_manage"（官方 / 预装 / 市场技能永不可触）；非 pinned；闲置超过 idle_days（默认 90 天）；目录真实存在。
The maximum action is reversible: MOVE to skills/.archive/<name>__<ts>/ plus deregistration — fully reversible; dry-run by default, --apply actually moves files. A report-only overlap / bloat scan also ships in scripts/skill_curator.py.最大动作 = MOVE 到 skills/.archive/<name>__<ts>/ + 注销注册，完全可逆；默认 dry-run，--apply 才动文件系统。另有 report-only 的重叠 / 膨胀扫描（scripts/skill_curator.py）。
The only "scheduler" in the system: a background review process with a 6-hour cooldown, triggered on session idle, never resident.系统里唯一的"调度器"：6 小时冷却的后台复查进程，会话空闲触发，非常驻。
The full pass carries three stages in order — Phase 5 wiki lint → M2 daily distill → M3 weekly promotion — L5, then L4, then L6. Each stage runs its own due-check: when not due, its overhead is near zero. Each stage fails silent: it never drags down the main flow.full pass 依次搭载三个环节：Phase 5 wiki lint → M2 daily distill → M3 weekly promotion（即 L5 → L4 → L6）。每个环节自带 due-check，not-due 时开销近似为零；每个环节 fail-silent，绝不拖垮主流程。
Every step is observable end-to-end: write journals, replayable correction chains, a jsonl log family, and telemetry events.全链可观测：写入日志、可回放的修正链、jsonl 日志族与遥测事件。
write_journal / ledger — every memory write is recorded.write_journal / ledger — 每次 memory 写入记录。
supersede chain — the full correction history of a fact is preserved (invalid_at soft invalidation) and replayable.supersede 链 — 事实修正历史全保留（invalid_at 软失效），可回放。
jsonl log family — distill.jsonl / promotions.jsonl / conflicts.jsonl under logs/wiki/.jsonl 日志族 — distill.jsonl / promotions.jsonl / conflicts.jsonl（logs/wiki/）。
telemetry_hooks — memory.flush / memory.distill / memory.promotion events.telemetry_hooks — memory.flush / memory.distill / memory.promotion 事件。
Forensic comparison against OpenClaw and Hermes (2026-08-12).对 OpenClaw / Hermes 的取证对比（2026-08-12）。
vs OpenClaw Dreaming (light / REM / deep three stages + Dream Diary): our distill + promotion are the equivalent of its consolidation + promotion, and our fact-level supersede audit chain is stronger — OpenClaw's LanceDB forget is an unauditable hard delete.vs OpenClaw Dreaming（light / REM / deep 三阶段 + Dream Diary）：我们的 distill + promotion 等价其巩固 + 晋升，且事实级 supersede 审计链更强——OpenClaw 的 LanceDB forget 是不可审计硬删。
vs Hermes Curator (pin / archive / consolidate / patch, idle-triggered): our Curator is pure rule-based archiving and lacks consolidate / patch actions (skill volume is small, not yet needed); Hermes' learning_graph (memory × skill visualization) and next-round prefetch are two things we have not built.vs Hermes Curator（pin / archive / consolidate / patch，空闲触发）：我们的 Curator 是纯规则归档，缺 consolidate / patch 动作（技能量小，暂不需要）；Hermes 的 learning_graph（记忆 × 技能可视化图）与下轮 prefetch 是我们尚未做的两项。
What is uniquely ours我们独有 — the six-way routing discipline; the supersede audit chain plus reinforce counting; and user-profile / self-model synthesis.六路路由纪律；supersede 审计链 + reinforce 计数；user-profile / self-model 合成。
| Layer层 | Trigger触发时机 | Frequency频率 | Budget预算 | Audit artifact审计产物 |
|---|---|---|---|---|
| L1 write | agent, in-sessionagent 会话内主动 | realtime实时 | MEMORY 3000 / USER 2000 chars; ≤10 topicsMEMORY 3000 / USER 2000 字符；≤10 topics | write_journal / ledgerwrite_journal / ledger |
| L2 extract | turn endturn 结束 | per-turn每轮 | low cost, no main model低成本，不动主模型 | —— |
| L3 flush | session idle会话空闲 | per idle session每次空闲会话 | ≤2 memory + ≤5 topic≤2 memory + ≤5 topic | telemetry memory.flush遥测 memory.flush |
| L4 distill | bg_review full passbg_review full pass | daily, ≤1×/day每日，最多 1 次 | ≤5 topic regens≤5 topic regen | distill.jsonl · memory.distilldistill.jsonl · memory.distill |
| L5 lint | bg_review full passbg_review full pass | every 7 days每 7 天 | rules first; 180s timeout规则先行；180s 熔断 | conflicts.jsonlconflicts.jsonl |
| L6 promote | bg_review full passbg_review full pass | weekly每周 | ≤8 candidates; ≤2 writes; ≤220 chars≤8 候选；≤2 写入；≤220 字符 | promotions.jsonl · memory.promotionpromotions.jsonl · memory.promotion |
| L7 curator | idle / report空闲 / 报告 | idle_days (90 d)idle_days（90 天） | 4-condition safety model4 条件安全模型 | archive move (reversible)归档移动（可逆） |
| L8 bus | session idle会话空闲 | 6 h cooldown6 小时冷却 | per-stage due-check各环节 due-check | —— |
| L9 audit | throughout贯穿全链 | continuous持续 | —— | supersede chain · jsonl familysupersede 链 · jsonl 日志族 |