@kenz1117/dsh-engram 0.7.13 → 0.7.15

package/README.en.md

@@ -66,11 +66,12 @@ Sensory and emotional dimensions (smell, temperature, emotional weight) are deli
 - **Retrieval-practice loop** (spaced repetition): `engram_review_queue` returns palace coordinates and placard cues but **never the body text**, forcing the model to recall first; `engram_review` reveals the entry and `engram_report grade` (0-5) self-rates it, advancing an SM-2 schedule (1 → 6 → round(prev × ease) days, reset on failure, ease floor 1.3). Entries under review scheduling **no longer take part in automatic decay** — their fate is decided by recall. Session-start injection reports how many memories are due.
 - **History backfill** (v0.7.3+): distils dsh's persisted past sessions into the palace, turn by turn — each session is written into the project store matching **its own cwd** (no cross-project bleed), reusing the live-capture throttling, redaction, and echo suppression, with a `(session, turn)` idempotency key so an interrupted run resumes. Backfilled entries never enter the due-today queue. **You choose the import rules** (time window / turns per session / total turn budget / auxiliary model / include subagent·seeded·no-cwd sessions) and see a zero-cost estimate (no LLM calls) before running; staging happens in a dedicated **"Backfill" tab** with progress (including a **breakdown of skip reasons**) and pause. Auxiliary calls **reuse the model you are currently using** by default (a past session's log records the provider/model of that time, which may no longer be available here), or you can pick one explicitly from the host's registered providers/models in the tab's "Auxiliary model" dropdown.
 - **Knowledge flywheel**: capture/save → contradiction candidates (high-similarity neighbors create `contradicts` edges on write, for the model/user to adjudicate) → hit reinforcement (confidence +0.05) → distillation (topic clusters merge into higher-level rules, supersedes chains, confidence inheritance) → decay (low-importance, long-unaccessed entries archive; restorable).
+- **Automatic belief consolidation** (v0.7.14+, off by default; yml `reflect: suggest|auto`): after a session ends, raw memories that have not yet served as evidence for any belief are consolidated by an auxiliary LLM into one-sentence **observations**. Unlike distillation, the raw memories are all kept — the observation is an **additional abstraction layer**: each belief carries its evidence-id list plus a proof count, and same-topic beliefs merge automatically by vector proximity (cosine ≥0.9), so one belief is **continuously refined** instead of duplicated. When new evidence lands near an existing belief (≥0.86) the belief turns `stale`, and the next consolidation feeds its raw evidence back to the model for a freshness review with four outcomes — `form` (new), `refine`, `still` (holds), `refute` (kept with its evidence chain for audit). In `auto` mode consolidation runs in the background after the session (5s timeout; failures never affect the conversation); in `suggest` mode it only enqueues a pending-consolidation task that the `engram_reflect` tool drains (the tool can also be invoked manually in any mode). The top 3 active beliefs by proof count are injected with the profile packet; stale/refuted beliefs are never injected. Schema upgraded to v12 (the observations table; incremental migration, zero data movement).
 - **Jev System-One adjudication** (v0.7.12+, off by default, the `jev` block in yml): the DEFER fuzzy-band three-way ruling (merge / accept / defer) and contradiction-edge confirmation can be delegated to an external Jev model; when Jev is unavailable (network failure / timeout / missing key) it silently falls back to the pure-rule four-way behavior, **never blocking writes**. Every adjudication leaves a side observation (verdict, whether Jev took part, elapsed time, fallback reason; up to 50 in-process, cleared on restart), so the panel's "Jev" view answers "why was this one deferred" at a glance. Five yml fields — the switch / API key / endpoint / model / timeout — can be overridden from the panel (stored at `<dbDir>/jev-config.json`, 0600 file inside a 0700 directory) and take effect on save without a restart; a "Test connection" button verifies endpoint and model reachability with the current input (unsaved values are tested as typed, never written to the override file); the key is returned only masked by GET, never leaves the process, and is never back-filled into the input. The three thresholds stay yml-only advanced settings, shown read-only.
 - **Entity lexicon** (v0.7.10+): the auxiliary LLM output of automatic capture and `engram_save` also extracts entity mentions (people/projects/tools/concepts, with optional aliases), resolved by normalized name against existing entities' names/aliases — a hit reuses the entity and refreshes its "last mentioned" stamp, a miss creates one; memories and entities link many-to-many, stored in the same store as the source memory. `engram_search` rows carry entity tags of the hits; entity-resolution failures silently skip the links and never block the memory write itself. The panel's "Entities" view filters by kind, searches names/aliases, and opens one entity to see every memory it pulls in.
 - **Automatic capture** (when `ingest` is enabled): each new turn's first step extracts candidate facts from the previous turn out of the session log, and session end captures the final turn too (failures leave a pending key that the next session replays; capture is idempotent per session+turn), written with low confidence and, when embeddings are available, disposed by the four-state rule toward near neighbors (merge-and-reinforce restatements / edge-and-defer suspected contradictions / write fresh) — memories accumulate without you saying "remember this".
 - **Provenance audit**: every memory records its source session, turn, and event seq; `engram_review` traces the full source chain, supersede chain, contradictions, and operation log; all writes/edits/forgets/distills/decays land in the operation log table.
-- **Web management panel** (v0.7.0+): the "Memory Library" tab in Settings has eight views — Today (a summary bar: memories / open / clarity + last-7-days counts + due today + health ring, followed by tour proposal, room directory, refurb list, health breakdown; a profile card in the side column shows the curated profile's current content and version-diff history), Palace (filters incl. due today, tour-route order, batch actions, list & edit), Corridor (corridor bird's-eye + retrieval bench), Log (last-7-days counts + the full merged op_log, filterable by operation class), Backfill, Episodes (episodes grouped by session, with anchor-neighborhood expansion), Entities (the entity lexicon: kind filters + name/alias search + per-entity linked memories), and Jev (System-One adjudication: config overrides + connection test + recent adjudications). A single 3-pill scope switcher in the header drives every panel; UI copy is bilingual zh/en and follows the host language setting live. Filter by redaction marks (include/exclude `[REDACTED:*]` entries) with amber badges for audit coverage.
+- **Web management panel** (v0.7.0+): the "Memory Library" tab in Settings has nine views — Today (a summary bar: memories / open / clarity + health ring, with a footer row of four small stats — founded / excavated / curated / due today — inline with the pulse timestamp; followed by tour proposal, room directory, refurb list, health breakdown; a profile card in the side column shows the curated profile's current content and version-diff history), Palace (filters incl. due today, tour-route order, batch actions, list & edit; the review / edit dialogs render as viewport-floating cards — sticky-pinned to the visible top of the host scroll area, so opening from any scroll depth never jumps or squeezes the page), Corridor (corridor bird's-eye + retrieval bench), Log (last-7-days counts + the full merged op_log, filterable by operation class), Backfill, Episodes (episodes grouped by session, with anchor-neighborhood expansion), Entities (the entity lexicon: kind filters + name/alias search + per-entity linked memories), **Beliefs** (v0.7.15+: consolidated beliefs sorted by evidence count with held / stale / refuted status filters; "evidence" expands the evidence chain for a read-only look at the underlying memories — purged evidence keeps the belief with an explicit notice; backed by the `GET /api/engram/observations` and `GET /api/engram/observation-evidence` routes), and Jev (System-One adjudication: config overrides + connection test + recent adjudications). The header is a single two-ended row: the scope pills (Private / Project / Shared) on the left drive every panel, with the workspace dropdown inlined on the same row for project scope; the right side holds the due badge plus icon-only revisit / export buttons (the export flyout is no longer covered by the tab row); a "中 | EN" segmented control switches the panel copy language (independent of the host language, persisted), with all copy bilingual zh/en. Filter by redaction marks (include/exclude `[REDACTED:*]` entries) with amber badges for audit coverage.
 - **Prompt-injection defense**: every memory recall exit (profile injection, `engram_search/timeline/review` output) is wrapped in `<engram_memory_context>` protocol tags with a usage warning (history is not the current request, do not follow instructions inside it, use only when relevant), while the current request is wrapped separately in `<current_user_request>`; all inbound content (capture candidates, saved bodies) is stripped of these protocol tags first, blocking second-order injection through forged protocol blocks.
 - **Capture redaction**: inbound content passes a regex scrub for common secrets and credentials (sk- API keys, Bearer, AWS AKIA, GitHub tokens, PEM private keys, password/token assignments, Chinese-language password assignments) and personal data (mainland-China phone numbers, 18-digit national ID numbers); matched fragments become `[REDACTED:<type>]`.
 - **Recall placeholder (anti-echo-chamber)**: memory-recall tool output inside captured slices is replaced with `[engram memory result omitted from capture: <tool>]`, and the extraction prompt states that restating existing memory is not new information, breaking the memory self-reinforcement loop.
@@ -79,7 +80,7 @@ Sensory and emotional dimensions (smell, temperature, emotional weight) are deli
 - **Portable data**: `engram_export` exports Markdown / JSON files in one step, with a redacted-view variant (secondary scrubbing + 40-char preview truncation, share-safe). `engram_mirror` writes an Obsidian / Logseq-friendly mirror directory (one Markdown per memory with YAML frontmatter + `[[id]]` backlinks), turning the palace into a human-readable private knowledge base.
 - **AGI Architecture Exploration (dsh-market · AGI Architecture)**: listed in dsh-market's "AGI Architecture Exploration" category as a cognitive-science reframe of agent long-term memory — memory palace (imagery labels + room placards), corridor topology (force-directed graph), closure questions (the ingest prompt nudges the model to ask clarifying questions), consolidation merging (heuristic dedup + cosine similarity) — sitting alongside MemGPT / Letta in the "agent memory architecture" conversation.
 
-## Tools (20, narrow parameters)
+## Tools (21, narrow parameters)
 
 | Tool | Purpose |
 |---|---|
@@ -102,6 +103,7 @@ Sensory and emotional dimensions (smell, temperature, emotional weight) are deli
 | `engram_ingest_history` | History backfill: distil past dsh sessions into the palace turn by turn (per-cwd stores; already-captured turns skipped). After each session's turns are captured, an auxiliary LLM generates a one-sentence session summary stored for timeline group headers (live sessions get the same summary at dispose once the final turn is captured; failures stay silent and never block). `dryRun` defaults to true (estimate only); pass `dryRun=false` to run. Use the Settings "Backfill" tab for large batches |
 | `engram_export` | Export Markdown / JSON files (data portability); `redactedView: true` emits a redacted view (secondary scrubbing + 40-char preview truncation, share-safe) |
 | `engram_distill` | Distill: merge same-topic clusters into higher-level rules (LLM) |
+| `engram_reflect` | Belief consolidation: abstract unconsolidated raw memories into proof-backed higher-level beliefs (four actions form/refine/still/refute; automatic vector-near-neighbor merging; raw memories stay active). Also drains the pending-consolidation queue created by reflect=suggest |
 | `engram_profile_edit` | Edit the curated profile block: `view` shows the current content and version history; `edit` writes a new version (optimistic lock via `expectedVersion`, conflicts fail loud); `rollback` restores a historical version (appended as a new version; history is never rewritten). The curated profile is injected ahead of the derived profile; every edit lands in the operation log, and the profile card on the panel's Today view shows version diffs |
 
 ## Configuration
@@ -123,6 +125,7 @@ Optional configuration (cordis.yml):
     modelCacheDir: '~/.dsh/engram/models'  # embedding model cache directory
     hfEndpoint: 'https://huggingface.co'   # model download endpoint; set a mirror behind restricted networks
     ingest: 'off'                   # automatic capture: off | light (user messages only, ≤2/turn) | eager (assistant messages too, ≤5/turn)
+    reflect: 'off'                  # automatic belief consolidation: off | suggest (only enqueue on session end; drain via engram_reflect) | auto (consolidate in the background after the session); engram_reflect stays manually callable in every mode
     # provider and model must be given as a pair: aux-LLM route override for capture/distill (parsed from the session log by default)
     # provider: 'deepseek'
     # model: 'deepseek-v4-flash'
@@ -166,6 +169,7 @@ Session agent                              Host half (Node)
   ├─ engram_save / search / review … ──────▶ ├─ SQLite dual stores (user.db / project-<origin hash>.db)
   │                                          ├─ FTS5 keyword track + local vector track, RRF fusion + ranking boost
   ├─ engram_distill ───────────────────────▶ ├─ aux-LLM distillation (cluster merge → supersedes chain)
+  ├─ session end (reflect enabled) ────────▶ ├─ belief consolidation (unconsolidated evidence → observations: near-neighbor merge / stale review)
   │                                          └─ automatic capture: session log → candidate facts (incl. the final turn at session end; ingest on)
   └─ Settings "Memory Library" tab ◀──────── ─── loopback API /api/engram/* (writes verify loopback Origin)
 ```
@@ -234,7 +238,7 @@ Profile text changes as the memory store changes — changes only land at turn b
 The palace IA (position-as-index / fixed routes / retrieval practice) is the indexing and audit skeleton; it evolves toward AI-native memory tiers and a temporal fact graph in three phases. Actual releases increment by 0.0.1; the phase labels are planning codenames only:
 
 - **Phase 1 · Flywheel closure** (done, v0.7.8): four-state write disposition (ACCEPT new / MERGE reinforce-in-place / DROP low-entropy / DEFER contradiction-pending — `engram_save` and capture return it per item); graduated decaying injection budget (160 chars for the first entry, ×0.9 per subsequent entry, floor 24, all configurable); redaction extended to Chinese passwords, national ID numbers, and phone numbers; evidence-gate wrap-up reminder (inject a reminder when a turn ends with unassessed search batches).
-- **Phase 2 · Memory tiers**: editable profile (**done, v0.7.9**: `engram_profile_edit` tool with version-chain rollback and panel diff audit, following Letta's core-memory model); a dedicated timeline index for episode memories (**done, v0.7.9**: `engram_episode_timeline` tool with date ranges, per-session grouping, temporal-neighborhood expansion, and a dedicated episode index; the panel "Episodes" view plus capture-time and live-session summary group headers landed in the same batch); automated distillation (cluster size + similarity thresholds, suggest/auto modes); configurable room capacity and placard scoring (defaults unchanged).
+- **Phase 2 · Memory tiers**: editable profile (**done, v0.7.9**: `engram_profile_edit` tool with version-chain rollback and panel diff audit, following Letta's core-memory model); a dedicated timeline index for episode memories (**done, v0.7.9**: `engram_episode_timeline` tool with date ranges, per-session grouping, temporal-neighborhood expansion, and a dedicated episode index; the panel "Episodes" view plus capture-time and live-session summary group headers landed in the same batch); automated belief consolidation (**done, v0.7.14**: the observations table (schema v12) with evidence chains and proof counts; triggered on session end via `reflect: suggest|auto`; near-neighbor merging continuously refines one belief; stale freshness review with the four actions form/refine/still/refute; active beliefs are injected with the profile; manual `engram_distill` stays available); configurable room capacity and placard scoring (defaults unchanged).
 - **Phase 3 · Temporal fact graph**: entity table with write-time extraction and resolution (**done, v0.7.10**: the entities + node_entities tables; capture/save auxiliary output carries entity mentions resolved on write; `engram_search` hits carry entity tags; panel "Entities" view + entity loopback routes); fact-level triples carrying `valid_at` / `invalid_at` windows with soft-invalidated conflicts (**done, v0.7.11**: the facts table (entity + time window + `replaced_by` + source memory), capture auxiliary output carries facts matched into the table by normalized entity name, `replaces` declares the soft-invalidation chain, `engram_search` gains `asOf` point-in-time lookback, the `engram_facts` tool reads and writes fact chains, and the panel entity page shows fact-chain details); automatic supports / refines / related edge creation; a LoCoMo-zh evaluation subset in CI as a retrieval-quality regression gate (deterministic Recall@k / MRR metrics).
 
 ## Acknowledgements

package/README.md

@@ -66,11 +66,12 @@ dsh plugin --profile web add @kenz1117/dsh-engram
 - **检索练习闭环**（间隔重复）：`engram_review_queue` 只给宫殿坐标与门牌线索、**不给正文**，迫使模型先主动回忆；`engram_review` 揭示核对，`engram_report grade`（0-5）自评推进 SM-2 调度（1 → 6 → round(prev × ease) 天，失败重置，ease 下限 1.3）。进入复习调度的条目**不再参与自动衰减**——命运由回忆结果决定。会话开始注入会提示今日待回忆条数。
 - **历史会话回填**（v0.7.3+）：把 dsh 已持久化的历史会话逐轮提炼进宫殿——默认按**每个会话自己的 cwd** 写进对应项目库（不串库），同样逐条判作用域（跨项目通用的个人偏好落私人库），复用实时摄取的节流/脱敏/防回声，靠 (会话, 轮次) 幂等键支持中断续跑（键固定在私人库，与实时路径共用）；回填条目不进今日复习队列（避免一次性回填淹没「今日待回忆」）。**导入规则由你选**（时间窗 / 单会话轮数 / 总轮数上限 / 辅助模型 / 是否含子代理·种子·无 cwd 会话），先估算（零成本、不调 LLM）再执行；设置页有独立的**「回填」tab**，可看进度（含**跳过原因分布**）与暂停续做。辅助调用默认**复用你当前在用的模型**（历史日志里记的是当年的 provider/model，在当前环境可能已不可用），也可在面板「辅助模型」下拉里从宿主已注册的 provider/model 中直接指定。
 - **知识飞轮**：摄取/保存 → 矛盾候选（写入时高相似近邻建 `contradicts` 边并报告，模型/用户裁决）→ 命中强化（confidence +0.05）→ 蒸馏（同主题簇合并为高层规律、supersedes 取代链、置信度继承）→ 衰减（低重要性且长期未访问归档，可恢复）。
+- **自动信念巩固**（v0.7.14+，默认关闭，yml `reflect: suggest|auto`）：会话结束后把「尚未成为任何信念证据」的原始记忆交辅助 LLM 巩固为一句话**信念（observation）**——与蒸馏的关键区别是原始记忆全部保留，信念是**叠加的抽象层**：每条信念带证据 id 列表与证据数（proof count），同主题经向量近邻（余弦 ≥0.9）自动归并、**持续细化同一条**而非反复新增。新证据与既有信念近邻（≥0.86）时信念先转 `stale`，下次巩固把原始证据一并喂给模型复核，结果四选一——`form` 新建、`refine` 细化、`still` 维持、`refute` 否定（保留证据链供审计）。`auto` 档会话结束后台巩固（5 秒超时、异常不影响对话）；`suggest` 档只登记待巩固任务，由 `engram_reflect` 工具一键执行（任何档位都可手动调）。active 信念按证据数取前 3 条随画像包注入会话，stale/refuted 不注入。schema 升级到 v12（observations 表，增量迁移、存量数据零搬运）。
 - **Jev 系统一裁决**（v0.7.12+，默认关闭，yml `jev` 子配置）：DEFER 模糊带的三路判定（并入 / 放行 / 搁置）与矛盾边确认可交外部 Jev 模型判决；Jev 不可用（网络失败 / 超时 / 未配密钥）时静默回落纯规则四态，**不阻断写入**。每次裁决旁路留观测（判定结果、Jev 是否参与、耗时、回落原因，进程内最多 50 条、重启清空），面板「裁决」视图一眼排查「为什么这条被搁置」。yml 的开关 / 密钥 / 端点 / 模型 / 超时五项可在面板覆盖（存 `<dbDir>/jev-config.json`，0600 文件 + 0700 目录），保存即时生效无需重启；「测试连接」按钮用当前输入即时验证端点与模型可达性（未保存也按当前值测、不落覆盖文件）；密钥 GET 只回掩码、明文不出进程、输入框不回填。三阈值保持 yml 高级配置，面板只读展示。
 - **实体词典**（v0.7.10+）：自动摄取与 `engram_save` 的辅助 LLM 输出同时抽取实体提及（人物/项目/工具/概念，可带别名），按归一化名与既有实体的 name/aliases 精确匹配消解——命中复用并刷新「最近提及」，未命中新建；记忆与实体多对多关联，随来源记忆所在 scope 分库。`engram_search` 命中行附关联实体标签；实体消解故障只静默跳过关联，不影响记忆落库。面板「实体」视图按类别筛选、按名称/别名搜索，点开单实体看它牵出的所有记忆。
 - **自动摄取**（`ingest` 配置开启时）：新一轮第一步从会话日志提取上一轮的候选事实，会话结束时补摄取最后一轮（失败留 pending 键，下次会话自动补做，幂等不重复），低 confidence 写入，嵌入可用时按四态处置近邻（复述并入强化 / 疑似矛盾建边待裁决 / 全新写入）——不说"记住"也能攒记忆。**逐条判宫殿**：提炼时同步判定作用域，只跟当前项目/仓库有关的（技术选型、项目约定、架构决策）进当前会话 cwd 对应的项目库，跨项目通用的（个人偏好、习惯、本人经历）进私人库——偏好跟人走、约定跟仓库走。
 - **来源审计**：每条记忆记录来源会话、轮次与事件 seq，`engram_review` 完整回查来源链、取代链、矛盾与操作日志；全部写入/修改/遗忘/蒸馏/衰减入操作日志表。
-- **Web 管理面板**（v0.7.0+）：设置页「记忆库」tab 分八个视图——今日（速览条：记忆 / 开放 / 清晰度 + 近 7 天计数 + 今日到期 + 健康分环；下面是入殿导航、房间目录、待翻新、健康分构成；侧栏画像卡展示 curated 画像当前内容与版本 diff 历史）、陈展（筛选含今日到期 / 巡游路线序 / 批量 / 列表与编辑）、走廊（走廊鸟瞰 + 检索实验台）、日志（近 7 天计数 + 两库合并的完整 op_log，可按操作类别筛选）、回填、往事（按会话分组浏览 episode，锚点邻近扩展）、实体（实体词典：类别筛选 + 名称/别名搜索 + 单实体关联记忆）、裁决（Jev 系统一裁决：配置覆盖 + 测试连接 + 近期裁决记录）。Header 三宫格驱动全局 scope（私人 / 项目 / 共享），全部数据源同步；**项目 scope 下三宫格右侧显示当前项目宫殿所属工作区**（标题 + 路径，默认跟随 GUI 当前工作区），旁边的工作区下拉可固定到某个工作区或切回「跟随当前会话」。界面文案中英双语，跟随宿主语言设置实时切换。支持按脱敏标记筛选（仅看/排除含 `[REDACTED:*]` 的条目）并给命中条目挂琥珀色徽标，方便审计脱敏覆盖面。
+- **Web 管理面板**（v0.7.0+）：设置页「记忆库」tab 分九个视图——今日（速览条：记忆 / 开放 / 清晰度 + 健康分环，底行「落成 / 发掘 / 整理 / 今日到期」四项小统计与诊脉时间同行；下面是入殿导航、房间目录、待翻新、健康分构成；侧栏画像卡展示常驻画像（curated）当前内容与版本 diff 历史）、陈展（筛选含今日到期 / 巡游路线序 / 批量 / 列表与编辑；「参观 / 修缮」以视口浮层呈现——sticky 钉在宿主滚动区可见顶部，长列表任意滚动位置打开都不跳动不挤压）、走廊（走廊鸟瞰 + 检索实验台）、日志（近 7 天计数 + 两库合并的完整 op_log，可按操作类别筛选）、回填、往事（按会话分组浏览 episode，锚点邻近扩展）、实体（实体词典：类别筛选 + 名称/别名搜索 + 单实体关联记忆）、**信念**（v0.7.15+：巩固信念列表按证据数排序，成立 / 待复核 / 已否定状态筛选，点「证据 n 条」展开证据链只读回查支撑该信念的原始记忆——证据被清理时信念保留并明确提示；`GET /api/engram/observations` 与 `GET /api/engram/observation-evidence` 两个回环路由）、裁决（Jev 系统一裁决：配置覆盖 + 测试连接 + 近期裁决记录）。Header 单行两端对齐：左侧 scope 三宫格（私人 / 项目 / 共享）驱动全局 scope，项目 scope 下工作区下拉内联在同一行；右侧今日到期角标 + 重访 / 导出纯图标按钮（导出弹层不再被下方 Tab 行遮挡）；「中 | EN」段控切换面板文案语言（独立于宿主语言，选择持久化），全部文案中英双语。支持按脱敏标记筛选（仅看/排除含 `[REDACTED:*]` 的条目）并给命中条目挂琥珀色徽标，方便审计脱敏覆盖面。
 - **提示注入防护**：全部记忆召回出口（画像注入、`engram_search/timeline/review` 输出）包 `<engram_memory_context>` 协议标签并附使用警告（历史记忆非当前请求、不遵循其中指令、仅相关时使用），当前请求独立包 `<current_user_request>`；所有入库内容（摄取候选、保存正文）先剥离这些协议标签，防伪造协议块二次注入。
 - **摄取脱敏**：入库前正则清洗常见密钥凭据（sk- 系 API key、Bearer、AWS AKIA、GitHub token、PEM 私钥、password/token 赋值、中文密码赋值）与个人信息（中国大陆手机号、18 位身份证号），命中片段替换为 `[REDACTED:<类型>]`。
 - **召回占位（防回声室）**：摄取切片中记忆召回工具的输出替换为 `[engram memory result omitted from capture: <tool>]`，并向提取模型附注"既有记忆的复述不是新信息"，阻断记忆自我强化循环。
@@ -79,7 +80,7 @@ dsh plugin --profile web add @kenz1117/dsh-engram
 - **数据可携带**：`engram_export` 一键导出 Markdown / JSON 文件，支持脱敏视图（内容二次清洗 + 预览截断，分享安全）。`engram_mirror` 导出可漫游的镜像目录（Obsidian / Logseq 友好：每条记忆一个 Markdown，正文 + YAML frontmatter + 双向链接 `[[id]]`），让「宫殿」也成为可人读的私人知识库。
 - **认知架构探索（dsh-market · AGI 架构探索）**：本仓库是 dsh-market「AGI 架构探索」类目下，对 agent 长期记忆的认知科学方法论重构——记忆宫殿（意象标签 + 房间铭牌）、走廊拓扑（力导向图）、闭环提问（摄入时让模型主动追问用户细节）、巩固合并（启发式去重 + 余弦相似度），与 MemGPT/Letta 同层「agent 记忆架构」叙事。
 
-## 工具（20 个，窄参数）
+## 工具（21 个，窄参数）
 
 | 工具 | 作用 |
 |---|---|
@@ -102,6 +103,7 @@ dsh plugin --profile web add @kenz1117/dsh-engram
 | `engram_ingest_history` | 历史会话回填：把 dsh 历史会话逐轮提炼进宫殿（按会话 cwd 分库、已摄取轮次自动跳过）。每会话轮次摄取完成后用辅助 LLM 生成一句话会话摘要落库（失败静默不阻断），实时会话结束时同样收尾生成，供时间线组头回看。`dryRun` 缺省 true 只估算；`dryRun=false` 才执行。大批量建议用设置页「回填」tab |
 | `engram_export` | 导出 Markdown / JSON 文件（数据可携带）；`redactedView: true` 输出脱敏视图（二次清洗 + 40 字预览截断，可安全分享） |
 | `engram_distill` | 蒸馏：同主题簇合并为高层规律（LLM） |
+| `engram_reflect` | 信念巩固：把尚未巩固的原始记忆经辅助 LLM 抽象为带证据数的高层信念（form/refine/still/refute 四动作，向量近邻自动归并，原始记忆保留不归档）；同时是 reflect=suggest 待巩固登记的执行出口 |
 | `engram_profile_edit` | 画像 curated block 编辑：`view` 看当前内容与版本历史；`edit` 写入新版本（乐观锁 `expectedVersion`，冲突 loud 失败）；`rollback` 回滚到历史版本（作为新版本追加，历史永不改写）。curated 画像优先于派生画像注入；编辑记录入操作日志，面板今日页画像卡可查版本 diff |
 
 ## 配置
@@ -123,6 +125,7 @@ dsh plugin --profile web add @kenz1117/dsh-engram
     modelCacheDir: '~/.dsh/engram/models'  # 嵌入模型缓存目录
     hfEndpoint: 'https://huggingface.co'   # 模型下载端点，网络受限可配镜像
     ingest: 'off'                   # 自动摄取：off | light（仅用户消息，每轮≤2条）| eager（含助手消息，每轮≤5条）
+    reflect: 'off'                  # 自动信念巩固：off | suggest（会话结束只登记待巩固，engram_reflect 执行）| auto（会话结束后台巩固）；任何档位均可手动调 engram_reflect
     # provider 与 model 必须成对提供：摄取/蒸馏的辅助 LLM 路由覆盖（缺省从会话日志解析）
     # provider: 'deepseek'
     # model: 'deepseek-v4-flash'
@@ -166,6 +169,7 @@ dsh plugin --profile web add @kenz1117/dsh-engram
   ├─ engram_save / search / review … ──────▶ ├─ SQLite 双库（user.db / project-<origin hash>.db）
   │                                          ├─ FTS5 关键词道 + 本地向量道 RRF 融合 + 排序 boost
   ├─ engram_distill ───────────────────────▶ ├─ 辅助 LLM 蒸馏（簇合并 → supersedes 链）
+  ├─ 会话结束（reflect 开启）──────────────▶ ├─ 信念巩固（未巩固证据 → observations：近邻归并 / stale 复核）
   │                                          └─ 自动摄取：会话日志 → 候选事实（含会话结束的末轮，ingest 开启时）
   └─ 设置页「记忆库」tab ◀────────────────── ─── 回环 API /api/engram/*（写操作校验回环 Origin）
 ```
@@ -235,7 +239,7 @@ pnpm bundle
 宫殿 IA（位置当索引 / 固定路线 / 检索练习）是索引与审计骨架，按三阶段向 AI 原生的记忆分层与时序事实图谱演进；实际发版按 0.0.1 递增，阶段代号仅为规划标签：
 
 - **阶段一 · 飞轮闭环补强**（已完成，v0.7.8）：写入四态回报（ACCEPT 新增 / MERGE 并入强化 / DROP 低熵 / DEFER 矛盾待裁决，`engram_save` 与摄取逐条返回处置）；画像注入改分级递减预算（首条 160 字、逐条 ×0.9、下限 24，可配置）；脱敏扩充中文密码 / 身份证 / 手机号三类；证据门收尾提醒（轮次结束存在未判定 assess 批次时注入提醒）。
-- **阶段二 · 记忆分层**：画像可编辑（**已完成，v0.7.9**：`engram_profile_edit` 工具 + 版本链回滚 + 面板 diff 审计，Letta core-memory 路线）；episode 情景记忆独立时间线索引（**已完成，v0.7.9**：`engram_episode_timeline` 工具，日期范围、按会话分组、时间邻近扩展，episode 专用索引；面板「往事」视图 + 摄取期与实时会话摘要组头同批完成）；蒸馏自动化（簇规模 + 相似度阈值，suggest/auto 两档）；房间容量与门牌评分可配置化（默认与现状一致）。
+- **阶段二 · 记忆分层**：画像可编辑（**已完成，v0.7.9**：`engram_profile_edit` 工具 + 版本链回滚 + 面板 diff 审计，Letta core-memory 路线）；episode 情景记忆独立时间线索引（**已完成，v0.7.9**：`engram_episode_timeline` 工具，日期范围、按会话分组、时间邻近扩展，episode 专用索引；面板「往事」视图 + 摄取期与实时会话摘要组头同批完成）；信念自动巩固（**已完成，v0.7.14**：observations 表（schema v12）带证据链与证据数，会话结束按 `reflect: suggest|auto` 触发，近邻归并持续细化、stale 新鲜度复核四动作 form/refine/still/refute，active 信念随画像注入；手动 `engram_distill` 蒸馏保留）；房间容量与门牌评分可配置化（默认与现状一致）。
 - **阶段三 · 时序事实图谱**：实体表与写入期实体抽取消歧（**已完成，v0.7.10**：entities + node_entities 两表，摄取/保存的辅助 LLM 输出带 entities 提及并消解落库，`engram_search` 命中附实体标签，面板「实体」视图 + 实体回环路由）；事实级三元组携带 `valid_at` / `invalid_at` 时间窗，冲突事实软失效不删除（**已完成，v0.7.11**：facts 表（实体 + 时间窗 + `replaced_by` + 来源记忆），摄取辅助 LLM 输出带 facts 并按实体归一化匹配落表，`replaces` 声明软失效链，`engram_search` 加 `asOf` 时点回看，`engram_facts` 工具读写事实链，面板实体页事实链详情）；supports / refines / related 自动建边；LoCoMo-zh 评测集进 CI 作检索质量回归门（确定性 Recall@k / MRR 指标）。
 
 ## 致谢