@modusensus/dsh-mneme 0.7.10 → 0.7.11

package/README.en.md

@@ -0,0 +1,368 @@
+<p align="center"><strong>中文 | <a href="docs/SEMANTIC.md">English (Semantic)</a></strong></p>
+
+# dsh-mneme
+
+English | [简体中文](README.md)
+
+[![npm version](https://img.shields.io/npm/v/@modusensus/dsh-mneme?color=blue&label=npm)](https://www.npmjs.com/package/@modusensus/dsh-mneme)
+[![license](https://img.shields.io/badge/license-MIT-green)](LICENSE)
+[![Awesome](https://awesome-dsh-plugin.com/badge.svg)](https://github.com/awesome-dsh-plugin/awesome-dsh-plugin)
+[![tests](https://img.shields.io/badge/tests-450%20passed-success)](https://github.com/modusensus/dsh-mneme)
+
+> A cross-session memory plugin for DeepSeek Harness: it lets the Agent remember you, remember your projects, and organize memories automatically. **Mneme** (Μνήμη) — named after Mnemosyne, the Greek goddess of memory who presides over memory and dreams, just as autoDream consolidates memories in the background.
+
+`dsh-mneme` is a [DeepSeek Harness (DSH)](https://github.com/deepseek-ai/deepseek-harness) plugin that gives Agents persistent cross-session memory. It draws on Claude's **Dream mechanism** and the **autoDream** implementation ideas from cc-haha / Claude Code — it does not just **store** memories, it also **consolidates them automatically** (deduplication, merging, conflict adjudication, summary generation), so the memory store keeps getting more refined with use.
+
+## ✨ Features
+
+### Memory Storage (SQLite + Markdown Mirror)
+
+- **SQLite primary storage**: `~/.dsh/memory/memory.db`, built-in `node:sqlite`, zero native dependencies
+- **Markdown mirror**: `preferences.md` / `projects.md` / `decisions.md` / `history.md` / `summary.md` — human-readable and hand-editable (**manual edits take priority** when merged back into the store)
+- **4+1 memory types**: `preference` / `project` / `decision` / `history` / `summary`
+- **Mirror sync state machine (v0.3.6+)**: the mirror is strongly consistent with the primary store, with sync debt modeled via `generation` (desired round) / `applied_generation` (applied round)
+  - Business write operations **atomically increment** the desired generation **within their own transaction** — even a crash after COMMIT but before rendering recovers on restart from the durable debt, never silently skipped (v0.3.8)
+  - `generation` is incremented with atomic SQLite statements — zero loss under multi-process concurrency; a `CHECK` upper bound rejects negative values/overflow
+  - Per-type `committed / failed / pending` receipts; the health endpoint distinguishes `ok / degraded / unknown`
+  - State write failures are never silent: sync failures are logged and leave debt behind, converging automatically on restart
+
+### Model Tools (7)
+
+| Tool | Function |
+|------|------|
+| `memory_save` | Save a memory (automatic dedup and merge by title) |
+| `memory_search` | Full-text search (Chinese-substring friendly; vector semantic search can be enabled) |
+| `memory_list` | Paginated listing by type (`include_archived=true` to view archived items) |
+| `memory_update` | Modify an existing memory |
+| `memory_delete` | Delete a memory |
+| `memory_forget` | Suppress injection (down-weighted rather than deleted; recoverable) |
+| `memory_archive` | Archive/restore memories (v0.2.5; archived items are hidden from listing/search/injection/consolidation, `archived=false` restores them) |
+
+### Auto Injection + Session Summary
+
+- **Auto injection**: at the start of a new session, a memory digest is injected (the `summary` first, plus a few high-importance items)
+- **Session summary**: at `turn/end`, an LLM distills the preferences/decisions/lessons of the session and stores them automatically (plugin-injected context is filtered out to avoid contamination)
+
+### autoDream Automatic Memory Consolidation 🧠
+
+- **Trigger**: fires asynchronously and automatically once memory count > 10 or total characters > 5000 (never blocks writes)
+- **Decision-list consolidation**: the LLM outputs a decision list of `keep` / `merge` / `archive` / `conflict` / `update` decisions, which the server validates and then applies item by item
+  - `merge`: merges entries with similar topics, keeping the most information-complete one
+  - `archive`: archives outdated/redundant entries (recoverable, never physically deleted)
+  - `conflict`: adjudicates contradictory information — the winner is kept, the loser is archived with a provenance note appended
+  - `update` (v0.2.1): directly corrects outdated/incorrect content of a single memory (single id / must actually change / not `summary` / 24h protection / ≤2 per run)
+- **Failure tracking (v0.2.1)**: when the user corrects a memory, it is written to the `failure_memories` table (old value/new value), accumulating data for future self-evolution
+- **Summary generation**: after consolidation, a "memory store overview" (single instance) is generated, injected with priority in the next session
+- **Fail-safe**: illegal LLM output (unknown id / invalid action / cross-type merge / out-of-range importance) rejects the entire decision list — the memory store is never corrupted
+- **Adjudication audit**: every run writes to the `dream_runs` audit table (input snapshot sha256 digest + full input snapshot + decision list + per-id disposition + receipt), replayable offline; merge / conflict / update are applied idempotently — replays and concurrent duplicate runs have no cumulative side effects; `update` records a `_before` snapshot
+
+#### dreamMaxTokens Tuning Guide
+
+The default `4096` covers ordinary memory stores. When the **memory volume is large** (tens of thousands of characters or more), the decision list and summary may exceed the default budget; scale it up by size:
+
+| Memory store size | Recommended `dreamMaxTokens` |
+|-----------|----------------------|
+| Ordinary (<10k chars) | `4096` (default) |
+| Medium (10k–50k chars) | `65536` |
+| Large (>50k chars) | `131072` (cap) |
+
+> With **reasoning models** (e.g. DeepSeek-R1-like), the model may spend the entire budget on reasoning and return an empty body (the log shows `no json array in llm output`). In that case, setting `dreamReasoningEffort` to `low` suppresses reasoning overhead and leaves the budget for the body output; the sleep side has the corresponding `sleepReasoningEffort`. The default `none` omits the field entirely, fully honoring the model's own default — behavior identical to earlier versions.
+
+### Sleep Mode: System-Level Sleep 💤 (v0.4.0, opt-in)
+
+Upgrades autoDream's "passive threshold triggering" into "proactive scheduled maintenance + tiered compression". Once the system has been idle for `sleepIdleMinutes` minutes, deep maintenance runs automatically. **Off by default** (`sleepModeEnabled: false`); once enabled:
+
+- **Interruptible**: implemented with AbortController — user activity aborts the current cycle (`noteWrite` resets the idle timer + the abort signal)
+- **Serially safe**: sleep cycles go through the `service.enqueue` serial queue, strictly non-overlapping with autoDream; `minRefTimeMs` prevents memories recalled after the snapshot from being mistakenly demoted
+- **Four-phase deep maintenance**:
+  1. `conflict_resolution`: store-wide conflict resolution, strictness configurable in three levels (gentle 0.92 / normal 0.85 / aggressive 0.75)
+  2. `archival_demotion`: tiered by `last_accessed_at` — not recalled for 30 days → compressed into a summary (original text kept in `_full_content`, losslessly restorable); 90 days → fully archived
+  3. `pattern_discovery`: the LLM scans recent memories to distill patterns, producing `type=pattern` memories, with strict evidence validation to prevent fabrication
+  4. `relation_completion`: detects orphaned entities and completes implicit relations (co-occurrence `related_to` / project `part_of` / technology `depends_on`)
+- **Fail-safe**: each phase has its own try/catch — an LLM failure only skips the corresponding phase; without an LLM route, the pure-rule fallbacks (demotion/relations) still run
+- **Audit continuity**: sleep cycles write to `dream_runs` with `run_type='sleep'`, sharing the audit table with autoDream for traceability
+
+> See `docs/SLEEP.md` for configuration; see `docs/MIGRATION.md` for migration notes.
+
+### Web Memory Panel
+
+Official settings panel → "Memory Store Settings" → "Memory" tab: browse by type, full-text search; once vector search is enabled, the "Semantic" toggle becomes available for vector recall.
+
+### User Settings (Profile / Rules) and Custom Commands ⚙️
+
+Official settings panel → "Memory Store Settings" tab:
+
+- **User profile**: a free-text passage describing the user (role, background, preferences), **injected into the system prompt every turn** so the Agent always follows it
+- **Rules**: a list of behavioral rules the Agent must obey (e.g. "give the conclusion first"), likewise injected every turn
+- **Custom commands**: register slash commands (`/name`); when triggered, the user-defined instruction content is handed to the Agent. Commands persist to SQLite, are automatically registered into the DSH command table at startup, and adding/removing takes effect in real time
+
+> Profile and rules are injected through a separate `[用户设置]` (User Settings) block (higher priority than the memory store), and are injected even when memories are empty.
+
+### Vector Search (Semantic Search) 🔎
+
+An optional capability: connect an OpenAI-compatible embeddings API so search can hit memories that are **literally different but semantically close**.
+
+**Configuration**: official settings → "Memory Store Settings" → scroll to the bottom "Vector Search" section:
+
+| Field | Description |
+|------|------|
+| `启用向量搜索` (Enable vector search) | Master switch; once on, a "Semantic" toggle appears in the memory panel |
+| `API 地址 (Base URL)` | OpenAI-compatible endpoint, e.g. `https://api.openai.com/v1`; also supports SiliconFlow, Zhipu, local Ollama, etc. |
+| `API Key` | The key for the corresponding service |
+| `模型名` (Model name) | Embedding model, e.g. `text-embedding-3-small`, `text-embedding-v3`, `bge-m3`, etc. |
+
+After saving, click "重建索引" (Rebuild Index) to backfill vectors for existing memories in bulk (newly written memories are embedded automatically). Then enter a query in the memory panel and click "语义" (Semantic) to recall semantically related results via vectors; if the vector service is unavailable, it automatically falls back to full-text search.
+
+> ⚠️ The key is stored only in the local `user_settings` table of `~/.dsh/memory/memory.db`; it is never uploaded and never written into the code repository.
+> You need an embedding model, not a rerank model: e.g. Alibaba Cloud's `text-embedding-v3` works, while `qwen3-vl-rerank` is a rerank model (it does not go through `/embeddings`).
+
+### Semantic Enhancement 🧠
+
+Since v0.2, a **fully offline semantic memory engine** (local models + reranking + clustering):
+
+- **Local embedding**: three selectable backends — ONNX (`Xenova/bge-small-zh-v1.5`, offline) / Ollama / OpenAI-compatible; failures fall back level by level automatically, with keyword search as the last resort
+- **Rerank fine-ranking**: `Xenova/bge-reranker-base` cross-encodes recalled candidates for reranking, improving Top-K accuracy
+- **autoDream semantic enhancement**: clusters memory vectors (`clusterMemories`) to automatically discover topically similar / potentially contradictory memories, making consolidation more precise
+- **Search pipeline**: hybrid recall (keywords + vectors) → Rerank → Top-K
+
+Configuration only requires setting `embedProvider` in `cordis.patch.yml` (default `openai`, preserving v0.1 behavior; switch it to `local` for offline). Upgrading requires no data migration.
+
+### Entity-Structured Memory (Entity Gene) 🧬
+
+Since v0.3.0, a new **memory gene** layer: extracts **named entities**, **attributes with a timeline**, and **relations between entities** from memories, upgrading search from "literal keywords" to "precise recall by entity/attribute".
+
+- **Three tables**: `entities` / `entity_attrs` (`valid_until` snapshot-style timeline) / `entity_relations`; opening an older database creates the tables automatically — idempotent, zero migration cost
+- **Automatic extraction**: once `entityExtractionEnabled=true`, newly written memories fire-and-forget trigger LLM extraction (same-name entity dedup, attributes stored on the timeline, relations appended; failures never block writes)
+- **Entity search** (`searchMemories` prefix routing, `entitySearchEnabled` on by default):
+  - `entity:阿尔托` → memories precisely associated via attributes (`_score 1.0`) rank ahead of keyword mentions (`_score 0.7`)
+  - `attr:国籍=芬兰` → memories exactly matching that attribute value
+  - `attr:国籍` → **all** currently valid memories for that attribute key (empty-value contract)
+- **autoDream integration**: `update` decisions write a `supersedes` self-reference (the attribute version is superseded); `merge` decisions migrate the loser's attribute ownership to the keeper (void if the keeper already has a current value for the same key)
+
+> 📖 See also: [Entity-Structured Memory Design](docs/ENTITIES.md) · [Semantic Enhancement Architecture](docs/SEMANTIC.md) · [Local Model Deployment Guide](docs/LOCAL_MODEL.md) · [Upgrading from v0.1](docs/MIGRATION.md)
+
+### Memory Quality Filter 🧼 (v0.4.6, on by default)
+
+Before writing to the store, every memory receives a **heuristic quality score** (a pure function — no I/O, no shared state): meta-memory vocabulary (talking about the memory system itself), self-referential type tags, overly short content, high repetition, and near-duplication of recent memories all deduct points (0-100):
+
+- `score ≥ 60`: stored normally
+- `30 ≤ score < 60`: `quality_score` is persisted and injection ranking is down-weighted by `importance × quality/100` (degraded)
+- `score < 30`: archived and flagged `low_quality` — still recallable via explicit search, but **never auto-injected**
+
+`memoryQualityFilter.enabled` turns the feature off entirely; `archiveThreshold` / `degradeThreshold` / `minContentLength` are tunable.
+
+### LLM Usage Audit 📊 (v0.4.6, on by default)
+
+Every **background LLM call** (autoDream consolidation + summary, autoSummarize compression) is written to the `llm_audit_logs` table: `tokens` / `duration` / `status` / `source` (which trigger produced it). Failed calls are recorded as `status=error` and never block the feature itself; `retentionDays` (default 90) purges expired rows at startup. Two new read-only APIs:
+
+- `GET /api/dsh-mneme/semantic/llm-audit?page=&pageSize=&source=` — paginated query + filter by source
+- `GET /api/dsh-mneme/semantic/llm-audit/stats?days=` — budget aggregated by source over the last N days (tokens / calls / failures)
+
+> Read-only endpoints; like list/search/semantic, they remain open even after `apiToken` is set.
+
+## 🆕 Recent Release Highlights
+
+| Version | Highlights |
+|------|------|
+| **v0.4.2** | autoSummarize custom model: the `summarizeProvider`/`summarizeModel` config options let you independently designate a lightweight model (e.g. qwen3.6-plus) for session summaries, saving main-model tokens; 473 tests green |
+| **v0.4.0** | System-level Sleep Mode: idle-triggered four-phase deep maintenance (conflict resolution / archival demotion / pattern discovery / relation completion), interruptible, serially safe, fail-safe; tiered compression releases cold memories; 471 tests green |
+| **v0.3.9** | Fixed 4 FAILs from the third-party audit: CAS made atomic within the same transaction, mirror degraded-receipt passthrough, per-type physical terminal-state convergence, strict integer validation for generation and stabilized concurrent initialization |
+| **v0.3.8** | All 6 runtime-blocking findings from the audit peer re-review fixed: desired generation atomically incremented within the same transaction (the crash window no longer skips silently), sync failures not silent, atomic generation increments (zero loss across processes), per-type committed/failed/pending receipts, explicit unknown on read failure, generation upper-bound/negative CHECK |
+| **v0.3.7** | Startup race fix: vector rebuild failing after a restart following manual edits to the md mirror (backfill moved to after init readiness + scheduleEmbed readiness gate) |
+| **v0.3.6** | Mirror sync state machine: generation/applied_generation debt modeling, F-NEW-03 mirror health status, persistent dirty + recoverMirror at startup |
+| **v0.3.0** | Memory gene: entity/attribute/relation three tables + timeline + entity search + autoDream supersedes |
+
+## 🗺️ Evolution Roadmap
+
+| Version | Status | Theme | Description |
+|------|------|------|------|
+| v0.2.x | ✅ Done | Semantic enhancement + reflection updates | Local embedding/rerank/clustering, `failure_memories` failure tracking |
+| v0.3.0 | ✅ Done | Memory gene | entities/attrs/relations three tables + timeline + entity search |
+| v0.3.6–0.3.8 | ✅ Done | Mirror consistency + audit hardening | generation sync state machine, 6 audit-peer runtime-blocking fixes, 450 tests green |
+| v0.3.9 | ✅ Done | Audit hardening A/B/D/F | compareAndUpdate same-transaction atomicity, degraded receipts, per-type physical terminal state, integer fail-closed, stable concurrent initialization |
+| **v0.4.0** | ✅ Done | System-level Sleep Mode | Idle-triggered four-phase deep maintenance (conflict resolution / archival demotion / pattern discovery / relation completion), tiered compression, interruptible serial fail-safe; 471 tests green |
+| **v0.4.2** | ✅ Done | autoSummarize custom model | `summarizeProvider`/`summarizeModel` config options, letting you independently designate a lightweight model (e.g. qwen3.6-plus) for session summaries and save main-model tokens; 473 tests green |
+| **v0.4.3** | ✅ Done | autoDream large-memory fix | issue#9 B+A: `dreamMaxTokens` cap raised 32768→131072 + `dreamReasoningEffort`/`sleepReasoningEffort` reasoning toggles (`none` by default, main conversation unaffected); 478 tests green |
+| **v0.4.4** | ✅ Done | autoDream decision coverage fix | issue#9 plan C: sliding window `dreamMaxSnapshotSize` (default 200, truncated by updated_at descending) + implicit keep `dreamImplicitKeep` (default true) + coverage floor `dreamMinExplicitCoverage` (default 50%) + fixed decision schema; 487 tests green |
+| **v0.4.5** | ✅ Done | Epistemic trust + recall eval | Memory credibility grading `trustEpistemicWeighting` (observation>inferred>subjective: retrieval ranking favors high-credibility memories, injection tags `[verified]`, dream merge/conflict favors the more credible side; opt-in, off by default) + retrieval evaluation `evaluateRetrieval` persisted to `recall_evals` (`evalPersistTestResults` opt-in, off by default; production retrieval always goes through `recall_runs`, unconditionally isolated); 518 tests green |
+| **v0.4.6** | ✅ Done | 8 fixes (vector pipeline + injection/quality/audit) | Vector pipeline fixes (embedSingle adaptation / `autoReindexOnBoot` backfill of existing data / `vector_meta` metadata) + injection semantic recall `hybridInject` + same-title append `content_history` + injection length caps (300 per item / 1500 per block) + memory quality filter `memoryQualityFilter` + LLM usage audit `llmAudit` (table + instrumentation + read-only APIs); 553 tests green |
+| **v0.4.7** | ✅ Done | Idempotent schema migrations | When the same db is opened concurrently, the `PRAGMA table_info` check and ALTER are non-atomic and may repeat `ADD COLUMN`, failing with a duplicate column name; switched to an `addColumn` helper that swallows the race (try/catch), unifying all 12 migration sites |
+| **v0.5.0+** | 🚀 Long-term | Self-evolving memory | Interest drift tracking + cross-workspace memory sharing (pending DSH support) |
+
+> All new capabilities ship as **toggleable features** (enabled/disabled via configuration), conservatively on by default and never breaking existing behavior. The `failure_memories` table and the autoDream decision engine have already paved the way for future reflective growth.
+
+## 📦 Installation
+
+### Prerequisites
+
+- [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) (DSH)
+- Node 24+ (`node:sqlite`)
+
+### Installation Steps
+
+#### Option 1: npm install (recommended)
+
+dsh-mneme is a **bundle** (it declares a `dsh.bundle` manifest); installation activates it automatically, no manual configuration required:
+
+```bash
+# 1. 安装插件（自动注册 bundle 层）
+dsh plugin --profile web add @modusensus/dsh-mneme
+
+# 2. 重启
+dsh web
+```
+
+> For custom configuration (thresholds, delays, etc.), override the defaults under `id: dsh-mneme` in `~/.dsh/profiles/web/cordis.patch.yml` (see the configuration table below).
+
+#### Option 2: Install from source
+
+```bash
+git clone https://github.com/modusensus/dsh-mneme.git
+cd dsh-mneme
+dsh plugin --profile web add .
+dsh web
+```
+
+#### Custom configuration (optional)
+
+It works out of the box with the defaults. To adjust, override in `~/.dsh/profiles/web/cordis.patch.yml`:
+
+```yaml
+- id: dsh-mneme
+  name: '@modusensus/dsh-mneme'
+  config:
+    memoryDir: ~/.dsh/memory
+    autoInject: true
+    autoSummarize: true
+    maxInjectedItems: 5
+    importanceThreshold: 3
+    autoDream: true
+    dreamThresholdCount: 10
+    dreamThresholdChars: 5000
+    dreamDelayMs: 2000
+```
+
+## ⚙️ Configuration
+
+| Key | Default | Description |
+|----|--------|------|
+| `memoryDir` | `~/.dsh/memory` | Memory storage directory (SQLite + Markdown) |
+| `autoInject` | `true` | Automatically inject memories at session start |
+| `autoSummarize` | `true` | Automatically distill a summary at session end |
+| `summarizeProvider` / `summarizeModel` | empty | LLM route override for summaries (empty = use the current session model); a lightweight model is recommended to save main-model tokens |
+| `maxInjectedItems` | `5` | Maximum number of memories to inject |
+| `importanceThreshold` | `3` | Minimum importance for injection (1-5) |
+| `autoDream` | `true` | Automatic memory consolidation switch |
+| `dreamThresholdCount` | `10` | Memory count threshold that triggers consolidation |
+| `dreamThresholdChars` | `5000` | Total character threshold that triggers consolidation |
+| `dreamDelayMs` | `2000` | Asynchronous consolidation delay (debounce) |
+| `dreamProvider` / `dreamModel` | empty | LLM route fallback for dream (defaults to the agent's default model) |
+| `dreamMaxTokens` | `4096` | Maximum tokens per dream LLM call (cap 131072; increase for large memory stores — see the tuning guide below) |
+| `dreamReasoningEffort` | `none` | Reasoning-effort passthrough for the dream LLM: `low` / `medium` / `high` / `none` (`none` = omit the field and use the model default; set `low` when a reasoning model exhausts its budget on reasoning and produces an empty body) |
+| `apiToken` | empty | Optional API auth token; once set, write operations and key endpoints require `Authorization: Bearer <apiToken>` |
+| `embedProvider` | `openai` | Semantic backend: `openai` (default, v0.1-compatible) / `local` (ONNX offline) / `ollama` |
+| `localEmbedModel` | `Xenova/bge-small-zh-v1.5` | Local ONNX embedding model |
+| `localEmbedDimension` | `512` | Local embedding vector dimension |
+| `localEmbedDevice` | `cpu` | Local inference device: `cpu` / `gpu` |
+| `localEmbedBatchSize` | `8` | Local embedding batch size (1-64) |
+| `ollamaBaseUrl` | `http://localhost:11434` | Ollama service address |
+| `ollamaModel` | `nomic-embed-text` | Ollama embedding model |
+| `embedModelCacheDir` | empty | Model cache directory (empty = user-level `~/.dsh/mneme/models`) |
+| `embedModelMirror` | `https://hf-mirror.com` | Mirror source for model downloads |
+| `vectorSearchTopK` | `20` | Top-K returned by vector search |
+| `vectorSearchThreshold` | `0.65` | Vector search similarity threshold |
+| `hybridSearchVectorWeight` | `0.6` | Vector weight in hybrid search |
+| `hybridSearchKeywordWeight` | `0.4` | Keyword weight in hybrid search |
+| `rerankEnabled` | `false` | Whether to enable rerank fine-ranking (the local onnxruntime model loads only when explicitly enabled) |
+| `rerankProvider` | `none` | Rerank backend: `local` / `none` (default `none`) |
+| `rerankModel` | `Xenova/bge-reranker-base` | Rerank cross-encoding model |
+| `rerankBatchSize` | `8` | Rerank batch size |
+| `rerankMaxCandidates` | `30` | Maximum number of rerank candidates |
+| `rerankScoreThreshold` | `0.1` | Rerank score threshold (candidates below it are dropped) |
+| `reflectionUpdateEnabled` | `true` | Master switch for `update` decisions |
+| `reflectionFailureTracking` | `true` | Master switch for failure tracking |
+| `reflectionUpdateMaxPerRun` | `2` | Maximum `update` decisions per consolidation run |
+| `reflectionUpdateMinAgeHours` | `24` | Protection period for newly created memories (hours) |
+| `entityExtractionEnabled` | `false` | Master switch for entity extraction (v0.3.0; always available at the storage layer) |
+| `entityExtractionModel` | empty | Dedicated extraction model (empty = use the agent's default model) |
+| `entityExtractionMaxEntities` | `10` | Maximum entities per extraction |
+| `entityExtractionMaxAttrs` | `20` | Maximum attributes per entity |
+| `entitySearchEnabled` | `true` | Switch for `entity:` / `attr:` prefix search |
+| `trustEpistemicWeighting` | `false` | Memory credibility weighting (v0.4.5, opt-in, off by default): memories are graded by source as `observation` > `inferred` > `subjective`; when enabled, retrieval ranking favors high-credibility memories, injection tags observation entries `[verified]`, and dream merge/conflict favors the more credible side; when off, `epistemic_status` is only persisted on save and does not participate in behavior |
+| `evalPersistTestResults` | `false` | Retrieval evaluation persistence (v0.4.5, opt-in, off by default): when enabled, `evaluateRetrieval` writes precision/recall/mrr snapshots into `recall_evals`; when off it only returns them to the caller without persisting. Production `searchMemories` auditing always goes through `recall_runs`, unconditionally never touching `recall_evals` |
+| `autoReindexOnBoot` | `true` | When existing memories lack embeddings and vectors are configured, a delayed, rate-limited background backfill rebuild starts after boot (set to `false` for manual rebuild only) |
+| `hybridInject` | `true` | Semantic-recall-first injection (v0.4.6, Bug4): when `injectCandidates` receives a non-empty query, it first recalls candidates via the vector index semantically, then fills in/deduplicates with rule-based filtering; empty query / no vectors falls back to the old logic |
+| `memoryQualityFilter` | `{enabled:true, archiveThreshold:30, degradeThreshold:60, minContentLength:10}` | Memory quality filter (v0.4.6, on by default): heuristic 0-100 scoring before write; meta-memory vocabulary/self-reference/overly short/duplicate/near-duplicate content deduct points; ≥60 stored normally, 30-60 down-weighted (injection ranked by importance×quality/100), <30 archived and flagged `low_quality` (still recallable via explicit search, never auto-injected) |
+| `llmAudit` | `{enabled:true, retentionDays:90}` | LLM usage audit (v0.4.6, on by default): every background LLM call (autoDream/autoSummarize) writes `llm_audit_logs` (tokens/duration/status/source); failures are recorded as error without blocking; read-only APIs `/api/dsh-mneme/semantic/llm-audit` + `/llm-audit/stats` |
+
+> 🔐 **API security**: DSH has no built-in authentication and by default listens only on `127.0.0.1`. The plugin API is open by default (so the web panel works out of the box). For protection (e.g. when exposed to a LAN), set `apiToken` in the configuration: write operations (profile/rules/commands) and key endpoints (`vector-config`, `vector-reindex`) require `Authorization: Bearer <token>` (the frontend settings panel accepts the same token), while the read-only `list` / `search` / `semantic` endpoints remain open. The `apiKey` returned by `/api/dsh-mneme/vector-config` is masked (`sk-***…`), while the stored plaintext is kept for actual calls; the frontend sending back an empty or masked value means "do not change the key".
+
+## 🏗️ Architecture
+
+```
+┌─────────────────────────────────────────────────┐
+│  存储层：SQLite (archived/forgotten 状态)         │
+│         + Markdown 镜像（人工可编辑，双向同步）    │
+├─────────────────────────────────────────────────┤
+│  服务层：saveWithDedupe / injectCandidates        │
+│         / mergeHumanEdits / onWrite 钩子          │
+├─────────────────────────────────────────────────┤
+│  模型接口：7 个工具 + 自动注入 + 会话摘要          │
+├─────────────────────────────────────────────────┤
+│  autoDream：阈值调度 → LLM 决策清单               │
+│            → 校验（fail-safe）→ 应用 → 摘要       │
+├─────────────────────────────────────────────────┤
+│  Web 面板：设置面板内嵌 + 浏览/搜索（含向量）    │
+└─────────────────────────────────────────────────┘
+```
+
+**Source layout**:
+
+```
+src/
+├── store.js          # SQLite 存储（CRUD、搜索、归档/遗忘、schema 迁移）
+├── mirror.js         # Markdown 镜像（渲染/解析，人工优先）
+├── service.js        # 领域逻辑（去重合并、注入筛选、写入钩子）
+├── config.js         # schemastery 配置 schema
+├── tools.js          # 7 个模型工具（defineTool）
+├── inject.js         # systemPrompt.context 动态注入
+├── summarize.js      # 会话结束 LLM 摘要
+├── dream.js          # autoDream 调度 + runDream（LLM 决策 + 摘要）
+├── dream/decisions.js# 决策校验（fail-safe）+ 决策应用
+├── entities/extractor.js # 实体抽取器（v0.3.0：LLM JSON 抽取 + 去重 + fail-safe）
+├── embedding.js      # OpenAI 兼容 embeddings 客户端 + 向量检索
+├── api.js            # HTTP 路由（Web 面板数据通道）
+└── index.js          # 插件接线
+lib/
+├── client.js         # Web 面板（手写 ModuleLoader bundle）
+└── *.js              # src 的同步分发产物
+test/                 # 450 个 node:test 测试（含审计与三轴线压测不变量）
+scripts/              # e2e-dsh.js 端到端演示 · stress-dsh.js 三轴线压测 · sync-lib.js 同步
+```
+
+## 🧪 Development
+
+```bash
+cd dsh-mneme
+npm install        # 安装 peer 依赖（以 devDependencies 形式，用于本地测试）
+npm test           # 运行 450 个测试
+npm run stress     # 三轴线压测：长会话检索 / 冲突仲裁 / 多 Agent 并发（离线 mock LLM）
+npm run sync       # 把 src/ 同步到 lib/（发布时由 prepack 钩子自动执行）
+```
+
+> The stress test (`npm run stress`) covers three axes: **long-session retrieval** (Recall@k, stale-residual rate), **conflict adjudication** (a replayable adjudication set: audit snapshot hash + receipt + idempotent replay), and **multi-Agent concurrency** (lost updates, duplicate merges, transaction/crash recovery). Every autoDream run writes to the `dream_runs` audit table (input snapshot digest + decision list + per-id disposition + receipt), so silent errors can be pinpointed even when the pass rate is high.
+
+> `lib/` is the synced distribution artifact of `src/` (`npm run sync`); `lib/client.js` is the hand-written web panel source and is unaffected by the sync.
+
+## 📄 Design Documents
+
+> Design documents live in `docs/` at the repository root; the links point there via the `../docs/` relative path (they resolve correctly on GitHub when opened from this directory).
+
+- [Entity-Structured Memory Design](docs/ENTITIES.md)
+- [Semantic Enhancement Architecture](docs/SEMANTIC.md)
+- [Local Model Deployment Guide](docs/LOCAL_MODEL.md)
+- [Upgrading from v0.1](docs/MIGRATION.md)
+
+## 📜 License
+
+MIT