memorix 1.9.4 → 1.9.6

package/docs/AGENT_OPERATOR_PLAYBOOK.md

@@ -41,7 +41,7 @@ For the 1.4 release line, the visible product shape is:
 - `memorix_project_context` / `memorix context "..."` is the normal black-box entry for non-trivial coding work: it assembles a bounded task Workset instead of injecting a generic memory dump; `memorix resume "..."` is the explicit CLI continuation projection
 - Code State keeps local snapshots and freshness links; a healthy pre-existing local CodeGraph index can add a bounded semantic outline, but Memorix never initializes or synchronizes that external index itself
 - `memorix knowledge` is an explicit review path for source-backed Markdown knowledge and canonical workflows. Do not initialize a versioned workspace or apply a proposal unless the user asks for that managed artifact
-- integration surfaces are agent-specific: Claude Code, Codex, GitHub Copilot CLI, Antigravity, and Hermes receive plugin packages; OpenClaw receives a compatible bundle; Pi and Oh-my-Pi receive package entries; Gemini CLI receives an extension package; OpenCode receives a plugin file and skill; Grok Build receives native hooks and `AGENTS.md` guidance while keeping MCP host-owned; DeepSeek Harness receives an MCP patch row, AGENTS.md guidance, and skills; Cursor and other agents receive MCP/rules/hooks where supported
+- integration surfaces are agent-specific: Claude Code, Codex, GitHub Copilot CLI, Antigravity, and Hermes receive plugin packages; OpenClaw receives a compatible bundle; Pi and Oh-my-Pi receive package entries; Gemini CLI receives an extension package; OpenCode receives a plugin file and skill; Grok Build receives native hooks and `AGENTS.md` guidance while keeping MCP host-owned by default (`--mcp http` writes Grok `config.toml`); DeepSeek Harness receives an MCP patch row, AGENTS.md guidance, and skills; Cursor and other agents receive MCP/rules/hooks where supported
 - privacy-safe diagnostics and receipts avoid raw chat, memory text, query text, tool payloads, and local file paths
 
 ---
@@ -257,7 +257,7 @@ What this installs depends on the target agent:
 - Pi: user Pi package with extension-based hook capture and a Memorix skill, registered through `pi install <path> --approve`; project-local setup uses `-l`.
 - Gemini CLI: local extension package under `~/.gemini/extensions/memorix`, extension-bundled stdio MCP, `GEMINI.md` context, hooks, commands, and skills. Antigravity CLI has an official Gemini CLI migration path, but Gemini CLI remains a separate target.
 - OpenCode: local plugin file, `opencode.json` MCP config, OpenCode skill, plus `AGENTS.md` guidance.
-- Grok Build: native hooks in `~/.grok/hooks/memorix.json` and guidance in `~/.grok/AGENTS.md`. Global hooks are trusted; project hooks require Grok `/hooks-trust` or `--trust`. Memorix does not write Grok's `config.toml` or MCP entries.
+- Grok Build: native hooks in `~/.grok/hooks/memorix.json` and guidance in `~/.grok/AGENTS.md`. Global hooks are trusted; project hooks require Grok `/hooks-trust` or `--trust`. MCP stays host-owned unless you pass `--mcp http`, which writes `[mcp_servers.memorix]` in Grok `config.toml`.
 - Antigravity: native plugin under `~/.gemini/config/plugins/memorix` or `.agents/plugins/memorix`, with `plugin.json`, `mcp_config.json`, `hooks.json`, rules, and skills.
 - OpenClaw: OpenClaw-compatible bundle under `~/.openclaw/extensions/memorix`, bundled `.mcp.json`, skills, and OpenClaw `HOOK.md`/`handler.ts` hook pack.
 - Hermes Agent: Hermes plugin under Hermes home (`%LOCALAPPDATA%\hermes` on native Windows, `~/.hermes` elsewhere, or `HERMES_HOME`), enabled in `config.yaml`, with plugin hooks, slash command, CLI command, skills, and MCP config.
@@ -404,7 +404,51 @@ memorix background ensure
 memorix background ensure 2>/dev/null
 ```
 
-Or use a launchd plist / systemd user service for true boot-time persistence.
+Or supervise `memorix serve-http` directly so a crash or reboot restarts the control plane.
+
+Linux systemd user unit (`~/.config/systemd/user/memorix-http.service`):
+
+```ini
+[Unit]
+Description=Memorix HTTP MCP control plane
+After=default.target
+
+[Service]
+Type=simple
+ExecStart=/usr/bin/env memorix serve-http --host 127.0.0.1 --port 3211
+Environment=PATH=%h/.local/bin:/usr/local/bin:/usr/bin
+Restart=on-failure
+RestartSec=3
+
+[Install]
+WantedBy=default.target
+```
+
+macOS LaunchAgent (`~/Library/LaunchAgents/com.memorix.http.plist`). Point `ProgramArguments` at `which memorix`:
+
+```xml
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+  <key>Label</key>
+  <string>com.memorix.http</string>
+  <key>ProgramArguments</key>
+  <array>
+    <string>/usr/local/bin/memorix</string>
+    <string>serve-http</string>
+    <string>--host</string>
+    <string>127.0.0.1</string>
+    <string>--port</string>
+    <string>3211</string>
+  </array>
+  <key>KeepAlive</key>
+  <true/>
+  <key>RunAtLoad</key>
+  <true/>
+</dict>
+</plist>
+```
 
 **Why this matters:** IDEs that use `serverUrl` (Windsurf, Cursor HTTP mode) connect to `http://localhost:3211/mcp` but cannot start the server. If the service is down, the IDE shows an MCP error with no recovery path. The user must run `memorix background start` or `ensure` manually.
 

package/docs/API_REFERENCE.md

@@ -81,6 +81,7 @@ memorix sync workspace --action scan
 memorix sync store status --json
 memorix sync store push
 memorix sync store pull
+memorix sync store push --scope user
 memorix sync store device rotate
 memorix ingest image --path ./diagram.png
 memorix media import --path ./architecture.png
@@ -92,7 +93,8 @@ memorix receipt --json --probe "release blocker"
 The CLI is for direct terminal use, not a 1:1 mirror of MCP tool names. It does not require an MCP connection. `--cwd` selects a Git project from any shell; an unbound terminal has project-visible access only, including transfer exports. `memorix identity join|use|clear` makes personal/team access and coordination explicit, while `--as <agent-id>` is the one-command equivalent for scripts. The only MCP-only area is the optional graph-compatibility tools (`create_entities`, `read_graph`, and related tools) for workflows that expect the official memory-server style graph API.
 
 `memorix sync store` is an explicit, opt-in multi-device observation relay. It
-uses the current Git project as its namespace, excludes personal/agent-targeted
+uses the current Git project as its namespace by default (`--scope project`),
+or every local project with `--scope user`. It excludes personal/agent-targeted
 and unqualified records, keeps a local retry outbox, and stores cursors locally.
 The default `status` action is preview-only. GitHub stores immutable JSONL event
 blobs; it does not receive `memorix.db`, `memorix.db-wal`, or `memorix.db-shm`.
@@ -153,6 +155,42 @@ SessionStart hooks keep the default minimal hint lightweight. When memory behavi
 
 The intended loop for agents is: get the project brief when it helps, inspect the suggested current files, use stale or unbound memory only as a lead, store durable outcomes after the work changes the project, and resolve obsolete memories.
 
+### Task Continuity
+
+Task Continuity is a small append-only ledger for one coding task. It keeps
+requirements, decisions, verification states, risks, and the final outcome
+together without copying the underlying observation or source evidence.
+
+CLI:
+
+```bash
+memorix continuity start --task "Harden the release path" --requirements "keep 1.9.x,run CI"
+memorix continuity record --task-id <id> --kind decision --content "Use the existing release gate"
+memorix continuity record --task-id <id> --kind verification --verification-status passed --content "Full CI passed"
+memorix continuity record --task-id <id> --kind verification --verification-id ci --verification-status passed --idempotency-key ci-pass-1 --content "Full CI passed"
+memorix context "continue the release path" --task-id <id>
+memorix continuity close --task-id <id> --status completed --content "Ready for release"
+```
+
+MCP uses `memorix_continuity` with actions `start`, `record`, `show`,
+`list`, and `close`. Pass the returned `taskId` to
+`memorix_project_context` to add a bounded `Task continuity` section to the Workset and
+receipt. Entries are project-scoped, sanitized for credentials, append-only, and
+bounded to 200 events per task.
+The ledger also exposes a deterministic outcome projection:
+`validated` means completion has passing verification, `at-risk` means a failure,
+block, or abandonment is present, `in-progress` means the task is still open, and
+`unverified` means it closed without enough passing verification. The numeric score is
+diagnostic-only; it is not used to silently reorder or rewrite memories.
+Use a stable `verificationId` when a verification obligation moves from `pending`
+to `passed`, `failed`, or `skipped`; without one, entries are treated as separate
+evidence. A completed task with passing checks and unresolved risks is reported as
+`validated-with-risks`. `idempotencyKey` makes retries return the original event
+instead of appending a duplicate. Event caps are enforced inside an immediate
+SQLite transaction.
+Continuity does not automatically promote anything to durable memory; verification
+and outcome signals remain explicit.
+
 ### Evidence Cards and Feedback
 
 Evidence Cards are the persisted provenance index for project memory. They do

package/docs/CONFIGURATION.md

@@ -102,6 +102,15 @@ Memorix resolves configuration in this order:
 Environment variables stay available for CI, MCP launchers, and temporary shell
 overrides. They are not the default user-facing setup path.
 
+For the HTTP control plane, `MEMORIX_HTTP_MAX_BODY_BYTES` bounds JSON request
+bodies at 10 MiB by default and accepts values up to 64 MiB. Keep the default
+unless a trusted local client needs larger payloads; this limit is not an
+authentication boundary. `MEMORIX_HTTP_AUTH_TOKEN` enables Bearer-token
+authentication for every endpoint except `/health`. Non-loopback binds are
+refused unless this token is set or the operator explicitly sets
+`MEMORIX_HTTP_ALLOW_UNAUTHENTICATED_BIND=1` behind a private network or
+authenticated reverse proxy.
+
 If you want the simplest setup, configure `~/.memorix/config.toml` once and stop
 there. Add `<git-root>/memorix.toml` only when a repository needs different
 models, memory behavior, or server defaults.
@@ -149,6 +158,49 @@ Memorix uses an explicit provider path as-is: for example,
 `https://ark.cn-beijing.volces.com/api/v3` remains `/api/v3`. Only a bare host
 without a versioned path receives the convenience suffix `/v1`.
 
+Atlas Cloud is also available as a memory LLM preset in `memorix init` and the
+configuration TUI:
+
+```toml
+[memory.llm]
+provider = "atlascloud"
+model = "deepseek-ai/deepseek-v3.2"
+```
+
+The preset uses `https://api.atlascloud.ai/v1` and the existing OpenAI-compatible
+chat completion transport. Set `ATLASCLOUD_API_KEY` or `MEMORIX_LLM_API_KEY` in
+your environment, or store `api_key` in your private global config. Do not put
+credentials in project TOML. Explicit `model` and `base_url` overrides still
+apply; model IDs must be available in your Atlas Cloud account. The Atlas preset
+does not fall back to OpenAI, Anthropic, or OpenRouter environment keys.
+
+This is opt-in background memory processing and may incur API charges. Setting
+`ATLASCLOUD_API_KEY` alone does not enable it. Embeddings and memcode's model
+registry are unchanged.
+
+OrcaRouter works through the generic OpenAI-compatible gateway path. Keep the
+provider label as `openai`, set its base URL, and use an explicit model ID:
+
+```toml
+[memory.llm]
+provider = "openai"
+base_url = "https://api.orcarouter.ai/v1"
+model = "openai/gpt-4o-mini"
+
+[agent]
+provider = "openai"
+base_url = "https://api.orcarouter.ai/v1"
+model = "openai/gpt-4o-mini"
+```
+
+Use `orcarouter/auto` only when per-request routing is intended. For memory
+formation, structured output, and embeddings, a fixed namespaced model is more
+reproducible. Set `MEMORIX_LLM_API_KEY` and, when using an embedding model
+through the same gateway, set `MEMORIX_EMBEDDING_API_KEY` independently.
+When the provider is unavailable, Memorix keeps the durable local write and
+falls back to its documented heuristic/BM25 paths; it does not treat a remote
+provider outage as a storage failure.
+
 ### `[embedding]`
 
 Used by semantic/vector search. This lane is intentionally separate from
@@ -397,10 +449,15 @@ Common keys:
 Store sync is disabled unless `MEMORIX_SYNC_PROVIDER` is explicitly set. It
 replicates approved project-visible observation events, never the live SQLite
 file or its `-wal` / `-shm` companions. The local SQLite store remains the
-canonical read/write store.
+canonical read/write store. Default scope is the current Git project. Pass
+`--scope user` to relay every local project through a separate `user-global`
+named per-user namespace. Set the same stable value on that user's devices;
+the value is required so operators sharing one remote cannot silently share a
+default namespace.
 
 ```text
 MEMORIX_SYNC_PROVIDER=fs|github|s3|postgres
+MEMORIX_SYNC_USER_NAMESPACE=your-stable-user-id   # required with --scope user
 ```
 
 The filesystem relay needs `MEMORIX_SYNC_FS_ROOT`. The GitHub relay needs a

package/docs/DEVELOPMENT.md

@@ -13,7 +13,7 @@ Memorix is a TypeScript project built around:
 
 ## Current Development Baseline
 
-The current release line is **1.9.4** and package metadata is kept synchronized
+The current release line is **1.9.6** and package metadata is kept synchronized
 with the root package and MCP Registry manifest.
 
 Contributors should assume the following areas are part of the current release line:

package/docs/DOCKER.md

@@ -29,6 +29,22 @@ Then open:
 - MCP endpoint: `http://localhost:3211/mcp`
 - health: `http://localhost:3211/health`
 
+The compose port is bound to `127.0.0.1` by default. Do not expose this
+unauthenticated control plane directly to the public internet. For a hosted
+deployment, put it behind an authenticated reverse proxy or a private network
+boundary before allowing remote clients to connect.
+
+Set `MEMORIX_HTTP_AUTH_TOKEN` in the service environment when the control plane
+must accept remote clients. Send it as `Authorization: Bearer <token>`; the
+`/health` endpoint remains unauthenticated for readiness checks. The supplied
+compose file uses an explicit unauthenticated-bind opt-in because its published
+host port is loopback-only.
+
+HTTP JSON bodies are limited to 10 MiB by default (dashboard maintenance
+requests use a 1 MiB sub-limit). Operators can set
+`MEMORIX_HTTP_MAX_BODY_BYTES`, up to 64 MiB, when a trusted local client needs
+larger media or tool payloads.
+
 Stop it with:
 
 ```bash
@@ -50,7 +66,9 @@ If you prefer `docker run`:
 
 ```bash
 docker build -t memorix:local .
-docker run --rm -p 3211:3211 -v memorix-data:/data memorix:local
+docker run --rm -p 127.0.0.1:3211:3211 \
+  -e MEMORIX_HTTP_ALLOW_UNAUTHENTICATED_BIND=1 \
+  -v memorix-data:/data memorix:local
 ```
 
 ---
@@ -60,9 +78,11 @@ docker run --rm -p 3211:3211 -v memorix-data:/data memorix:local
 `memorix serve-http` accepts a `--host` flag:
 
 - **Local** (default): binds to `127.0.0.1` — only accessible from the host machine
-- **Docker**: the Dockerfile sets `--host 0.0.0.0` — accessible from outside the container, which is required for port mapping (`-p 3211:3211`) to work
+- **Docker**: the Dockerfile sets `--host 0.0.0.0` — accessible from outside the container, which is required for port mapping (`-p 3211:3211`) to work. Set `MEMORIX_HTTP_AUTH_TOKEN`, or explicitly opt into an unauthenticated private-network bind with `MEMORIX_HTTP_ALLOW_UNAUTHENTICATED_BIND=1`.
 
-If you run `serve-http` manually inside a container, make sure to pass `--host 0.0.0.0` or the port will not be reachable from the host.
+If you run `serve-http` manually inside a container, make sure to pass
+`--host 0.0.0.0` plus one of those explicit security settings, or the server
+will refuse to start.
 
 ---
 

package/docs/INTEGRATIONS.md

@@ -53,7 +53,7 @@ Generated guidance also has scope:
 | CodeBuddy Code | `memorix setup --agent codebuddy --global` | CodeBuddy local marketplace plugin | `~/.codebuddy/memorix-local` marketplace with `.codebuddy-plugin`, bundled stdio MCP, skills, and hooks | Setup attempts `codebuddy plugin marketplace add` then user-scope install. It does not alter existing CodeBuddy settings, permissions, or model configuration; review third-party hooks in CodeBuddy's `/hooks` flow before enabling them. |
 | GitHub Copilot CLI | `memorix setup --agent copilot --global` | Copilot CLI plugin package | Local plugin under `~/.copilot/plugins/local/memorix` with MCP, skills, and hooks | Setup attempts `copilot plugin install <local-path>` when Copilot CLI is available. |
 | Cursor | `memorix setup --agent cursor --global` | Cursor MCP and rules config | Cursor MCP config, `.cursor/rules/memorix.mdc`, skills, and hook guidance | Reload Cursor after setup so it can pick up project config changes. |
-| Grok Build | `memorix setup --agent grok --global` | `~/.grok/hooks/memorix.json` + `~/.grok/AGENTS.md` | Lifecycle hooks on SessionStart, SessionEnd, UserPromptSubmit, PostToolUse, PostToolUseFailure, PreCompact, PostCompact, Stop, and StopFailure, plus Grok project/global guidance | Uses Grok's own JSON hook directory and never writes `~/.grok/config.toml` or MCP entries. Global hooks are trusted; project hooks require Grok `/hooks-trust`. Set `GROK_HOME` to use a custom Grok home. The command is `memorix hook --agent grok`. |
+| Grok Build | `memorix setup --agent grok --global` | `~/.grok/hooks/memorix.json` + `~/.grok/AGENTS.md` | Lifecycle hooks on SessionStart, SessionEnd, UserPromptSubmit, PostToolUse, PostToolUseFailure, PreCompact, PostCompact, Stop, and StopFailure, plus Grok project/global guidance | Uses Grok's own JSON hook directory. MCP stays host-owned unless you pass `--mcp http`, which writes `[mcp_servers.memorix]` with `url = "http://localhost:3211/mcp"` in Grok `config.toml`. Global hooks are trusted; project hooks require Grok `/hooks-trust`. Set `GROK_HOME` to use a custom Grok home. The command is `memorix hook --agent grok`. |
 | Gemini CLI | `memorix setup --agent gemini-cli --global` | Gemini CLI extension | Extension under `~/.gemini/extensions/memorix` with MCP, `GEMINI.md` context, hooks, commands, and skills | Antigravity CLI has an official Gemini CLI migration path, but Gemini CLI remains an active standalone Google CLI target. |
 | OpenCode | `memorix setup --agent opencode --global` | OpenCode local plugin file | Global setup writes `~/.config/opencode/plugins/memorix.js`, `~/.config/opencode/opencode.json`, skills, and `AGENTS.md` guidance; repo-local setup writes the `.opencode/` equivalents. | OpenCode loads local plugin and skill files from its config scope. |
 | OpenClaw | `memorix setup --agent openclaw --global` | OpenClaw compatible bundle | `~/.openclaw/extensions/memorix` with `.mcp.json`, skills, and OpenClaw `HOOK.md`/`handler.ts` hook pack | Setup attempts `openclaw plugins install <path> --force` and `openclaw hooks enable memorix` when OpenClaw is available. |

package/docs/KNOWN_ISSUES_AND_ROADMAP.md

@@ -1,6 +1,6 @@
 # Memorix 已知边界与路线图
 
-> 最后审阅：2026-08-14（对照 v1.4.3）
+> 最后审阅：2026-09-22（对照 v1.9.6 发布内容）
 
 这份文档说明仍然成立的产品边界、风险和方向；它不是发布流水账。
 
@@ -10,6 +10,66 @@
 
 ---
 
+## 1.9.6 基础设施收口
+
+> 本节记录随 v1.9.6 合并并发布的基础设施整改及其验收边界。版本号仍保持
+> 在 1.9.x，不跳跃到 1.10。
+
+### 已完成并通过回归
+
+- SQLite 数据库句柄采用 lease + 有上限的 LRU registry；HTTP session 有硬上限、
+  pending admission、过期回收和关闭释放；health 暴露 SQLite、session、runtime
+  与 event-loop 指标。
+- MCP modern bridge 按请求创建协议 server，按项目复用有界 business runtime；
+  legacy business store 入口有串行门，避免跨项目并发切换污染。
+- Task Continuity 采用 `requirement -> verification -> evidence -> outcome` 的
+  可更新验证项语义，支持 `validated-with-risks`、幂等键、状态替换和 immediate
+  transaction 上限保护。
+- Orchestrator 记录脱敏的 effect ledger 与结构化 trace，标记成功、失败、未知和
+  重放要求；Memory Quality Gate 覆盖命中、精度、隔离、过期拒绝、延迟、RSS、
+  token 预算和失败归因。
+- `npm run lint`、`npm run build`、`npm run gate:memory-quality`、全量 Vitest
+  与现代 MCP/HTTP 集成回归均通过；本轮没有待处理的发布阻塞缺口。
+
+### 明确保留的产品边界
+
+- SQLite 底层仍是同步 API；本轮通过 runtime 串行门、资源上限、event-loop 指标和
+  quality gate 控制风险，没有虚假宣称已经异步化。
+- 协议层 modern MCP 保持 stateless；官方 SDK 的每请求 server 生命周期与业务
+  runtime 复用是两个不同层次，Durable Tasks 仍按既有契约不支持。
+- Memory Quality Gate 是确定性本地基准和回归门，不把一次通过误报成所有模型、
+  所有项目规模下的检索质量保证。
+
+---
+
+## 1.9.5 历史发布结论与剩余后续项
+
+> `v1.9.5` 已发布，版本标签为 `658231f`；1.9.x 容量维护已通过 #306
+> 合入当前主线 `3a3dd49`。本节保留发布证据和仍有外部依赖的事项，不再把
+> 已完成的门槛伪装成未完成工作。
+
+### 发布门槛
+
+| 优先级 | 项目 | 当前结论 | 1.9.5 完成标准 |
+|---|---|---|---|
+| P1 | HTTP 控制面安全与生命周期 | 已完成并随 v1.9.5 发布。 | 回环默认、非回环显式安全策略、请求体上限、监听失败、后台清理、HTTP/SQLite 关闭和真实 smoke 均通过。 |
+| P1 | 项目隔离 | 已完成并随 v1.9.5 发布。 | 图谱实体/边按项目存储，Dashboard/MCP 查询、替换、删除和旧 JSONL 迁移测试通过。 |
+| P1 | 迁移失败语义 | 已完成并随 v1.9.5 发布。 | legacy observations/subdirectories 具备锁和原子写入；坏 JSON fail-closed，不伪装成空库。 |
+| P1 | Store / SQLite 生命周期 | 已完成；本轮容量切片补齐了缓存、查询和维护保留边界，没有已知发布阻塞。 | SDK 同目录引用计数、sync ID/tombstone、活锁保护、关闭路径、多实例与容量回归通过。 |
+| P2 | 内存与容量边界 | 已完成并通过 #306 合入主线；压力验证也纳入本轮回归。 | TeamStore SQL 查询上限、failed job 30 天保留、三类 embedding cache 字节预算、HTTP project/store cache 驱逐均已有代码和回归覆盖。 |
+| P1 | 发布契约 | 已完成并随 v1.9.5 发布。 | tag/version/commit 校验、npm 传播等待、check-only prepublish、npm 和 MCP Registry 发布均已验证。 |
+
+### 新 issue / PR 处置
+
+- [#301](https://github.com/AVIDS2/memorix/issues/301) / [#303](https://github.com/AVIDS2/memorix/pull/303)：已通过 #305 合并并随 v1.9.5 发布；原 PR 已标记为 superseded。
+- [#302](https://github.com/AVIDS2/memorix/issues/302) / [#304](https://github.com/AVIDS2/memorix/pull/304)：已通过 #305 合并并随 v1.9.5 发布；原 PR 已标记为 superseded。
+- [#300](https://github.com/AVIDS2/memorix/pull/300)：Atlas Cloud 是可选 provider；当前收口分支补齐了配置隔离和完整 CI，确认后合入 1.9.x，不改变默认 provider。
+- [#297](https://github.com/AVIDS2/memorix/issues/297)：外部 awesome-list 收录请求，需按对方仓库的贡献流程处理，不是 Memorix 代码阻塞。
+- [#283](https://github.com/AVIDS2/memorix/issues/283)：已完成文档收口；OrcaRouter 通过现有 OpenAI-compatible `base_url` 配置使用，不需要新 provider 枚举。
+- [#49](https://github.com/AVIDS2/memorix/issues/49) 与 [#3](https://github.com/AVIDS2/memorix/issues/3)：外部集成提案；当前没有可验证的宿主身份/Hook 契约，不纳入本轮维护代码。
+
+---
+
 ## 当前产品边界
 
 Memorix 是本地优先、可插拔的 Agent Memory 系统。它给不同 Agent 或同一 Agent 的新会话提供同一个项目记忆层，并把 CLI 作为最完整、最稳定的控制面；MCP 是面向 Agent 的受控入口，而不是唯一入口。

package/docs/MULTI-DEVICE-SYNC-RESEARCH.md

@@ -89,6 +89,10 @@ backup is not a merge input.
       version/audit record.
 - [x] Personal, agent-targeted, candidate, ephemeral, and other-project rows
       are excluded by default.
+- [x] `--scope user` is an explicit opt-in that relays every local project's
+      eligible rows through a separate `user-global` namespace. Privacy
+      filters still drop personal, agent-targeted, candidate, and ephemeral
+      rows. The live SQLite file is still never copied.
 - [x] Copying a data directory produces a clear device-clone error and a
       documented `device rotate` recovery path.
 - [x] GitHub relay protocol smoke proves only JSONL/snapshot artifacts are uploaded;

package/docs/PERFORMANCE.md

@@ -39,9 +39,17 @@ user is allowed to remember.
   budget is large and can be raised with `MEMORIX_EMBEDDING_CACHE_MAX_BYTES`,
   but it is deliberately separate from the durable memory corpus. Cache
   eviction never deletes a memory; a later vector backfill can regenerate it.
+- FastEmbed and Transformers caches use the same byte budget and reject cache
+  files above a bounded load size. A high-dimensional or unusually large
+  vector is skipped instead of exceeding the configured budget.
+- Coordination lists are SQL-limited: agents, locks, tasks, and inboxes return
+  at most 100 rows by default and 500 rows at most per request. Maintenance
+  history keeps completed jobs for 7 days and failed diagnostics for 30 days.
 - `serve-http` `/health` exposes `rss`, V8 heap, external memory, array-buffer
   memory, and the V8 heap limit so a real saturation report can distinguish
-  JavaScript retention from native/model memory.
+  JavaScript retention from native/model memory. It also exposes SQLite handle
+  cache/lease counts, modern runtime pool counts, HTTP session admission counts,
+  and event-loop lag percentiles.
 
 ## SQLite Storage Reliability Contract
 
@@ -129,13 +137,15 @@ On the release development machine used for this check, the healthy HTTP service
 | Knob | Default | Use When |
 | --- | --- | --- |
 | `MEMORIX_SESSION_TIMEOUT_MS` | `43200000` (12 h) | Set a shorter GC window for supervised clients, or `0` to disable idle session GC |
+| `MEMORIX_MAX_HTTP_SESSIONS` | `64` (max `512`) | Bound concurrent stateful HTTP MCP sessions; excess initialization requests receive `429` |
 | `MEMORIX_FORMATION_TIMEOUT_MS` | `12000` (12 s) | Raise when LLM-backed formation should outlive slow proxy/provider hops |
 | `MEMORIX_LLM_API_KEY` / `OPENAI_API_KEY` | unset | Enable LLM-backed enrichment, extraction, rerank, or skill generation |
 | `MEMORIX_LLM_TIMEOUT_MS` | `30000` (30 s) | Bound a single LLM-backed extraction/resolve call |
 | `MEMORIX_RERANK_TIMEOUT_MS` | 30000 | Bound HTTP and LLM rerank calls |
 | `MEMORIX_RERANK_PROVIDER` | `off` | Set `http` to enable optional HTTP rerank |
 | `MEMORIX_RERANK_BASE_URL` | `[memory.llm].base_url` | Compatible `/rerank` API root (path `/rerank` is appended) |
-| `MEMORIX_EMBEDDING_CACHE_MAX_BYTES` | `268435456` | Disposable in-process API-vector cache budget; raise for large warm caches, or lower when the host is memory constrained |
+| `MEMORIX_EMBEDDING_CACHE_MAX_BYTES` | `268435456` API / `67108864` local providers | Disposable embedding-vector cache budget; raise for large warm caches, or lower when the host is memory constrained |
+| Team coordination list APIs | `100` default / `500` maximum | Bound agent, lock, task, and inbox result sets; use task/agent detail endpoints for individual records |
 | `MEMORIX_ORAMA_HYDRATION_THRESHOLD` | `10000` | Above this durable corpus size, skip full Orama hydration and use persistent SQLite/semantic indexes; this is a working-set threshold, not a memory quota |
 | `MEMORIX_SEMANTIC_INDEX` | `auto` | `auto` uses the optional local LanceDB shadow index when installed; `off` or `orama` keeps the legacy semantic path |
 | `MEMORIX_SEMANTIC_INDEX_THRESHOLD` | `10000` | Minimum vectors before the optional HNSW/SQ index is trained; vectors remain searchable before training |
@@ -147,6 +157,19 @@ On the release development machine used for this check, the healthy HTTP service
 | `memorix retention archive` | explicit | Archive expired memories when the project gets noisy |
 | `memorix memory deduplicate` / `consolidate` | explicit | Reduce duplicate or scattered memory records |
 
+Modern stateless MCP creates a fresh protocol server per request as required by
+the official SDK. Memorix reuses the project business runtime behind that shell,
+with a bounded pool and a reference count. The pool key includes both the shared
+SQLite `dataDir` and the verified project root because the default storage layout
+is flat across projects. Before a pooled tool call, Memorix restores the project
+configuration and legacy store context; requests pass through a single runtime
+gate to preserve project isolation. Long-lived SSE GET streams bypass the gate.
+
+Orchestrator runs persist redacted effect records for observed tool calls. The
+ledger stores tool name, risk tier, replay policy, a fingerprint, and settlement
+status, never raw tool arguments or output. An effect marked `unknown` after a
+crash must be reconciled before a dangerous retry.
+
 ## Operator Guidance
 
 - For memory-only use, prefer stdio MCP or a lightweight `memorix_session_start`; do not join orchestration coordination state by default.
@@ -191,6 +214,27 @@ Results are machine-specific and should be compared on the same OS and Node
 version. The release gate is an integrity and regression check, not a public
 latency guarantee.
 
+### 1.9.x Capacity Closeout Evidence (2026-09-21)
+
+The 40,000-record gate passed on Windows with Node `v22.22.3`, SQLite-backed
+storage, and embeddings disabled after the capacity guardrails landed:
+
+| Measurement | Result |
+| --- | ---: |
+| Seed and index | 51.4 s |
+| Steady-state write p50 / p95 | 0.87 / 1.51 ms |
+| SDK lexical search | 25.4 ms |
+| HTTP ready | 2.20 s |
+| MCP initialize / bind | 160 / 344 ms |
+| MCP context / search / store | 768 / 428 / 556 ms |
+| Hook write | 3.75 s |
+| SDK reopen | 129.8 ms |
+| Service peak RSS | 408.7 MiB |
+| Durable records after MCP and hook writes | 40,002 |
+
+The planted record remained searchable, the HTTP/MCP and hook writes were
+durable, and the reopened SDK observed the project-visible records.
+
 ### 1.9.3 Release Evidence (2026-09-14)
 
 The 40,000-record gate passed on Windows with Node `v22.22.3`, SQLite-backed

package/docs/README.md

@@ -32,6 +32,7 @@ The public docs are organized by user intent:
 | --- | --- |
 | CLI commands and MCP tools | [API_REFERENCE.md](API_REFERENCE.md) |
 | Memory Autopilot, Code State, and context packs | [API_REFERENCE.md § Memory Autopilot, Code State, and Context Packs](API_REFERENCE.md#memory-autopilot-code-state-and-context-packs), [1.2 Code State](1.2.0-CODE-STATE.md), and [1.2 Workset Retrieval](1.2.0-WORKSET-RETRIEVAL.md) |
+| Requirement, decision, and verification continuity | [API_REFERENCE.md § Task Continuity](API_REFERENCE.md#task-continuity) |
 | Reviewed episodic, semantic, and procedural long-term memory | [1.3 Memory Architecture](1.3-MEMORY-ARCHITECTURE.md) |
 | Browse memory, maintenance, and read-only Coordination Status | [Dashboard](DASHBOARD.md) |
 | Reviewable knowledge pages and workflow inheritance | [API_REFERENCE.md § Knowledge Workspace and Workflows](API_REFERENCE.md#knowledge-workspace-and-workflows), [1.2 Knowledge Workspace](1.2.0-KNOWLEDGE-WORKSPACE.md), and [1.2 Workflow Inheritance](1.2.0-WORKFLOW-INHERITANCE.md) |
@@ -125,7 +126,7 @@ Historical/deep-reference documents may describe older designs. If they conflict
 
 ## Current Product Line
 
-The current product line is **1.9.4**. Multi-device store sync is opt-in and
+The current product line is **1.9.6**. Multi-device store sync is opt-in and
 keeps local SQLite canonical; existing users do not upload anything until they
 configure a relay provider explicitly.
 The authoritative acceptance contract for the current implementation is

package/docs/SETUP.md

@@ -151,7 +151,7 @@ memorix serve
 
 Use this when your agent only needs a raw local stdio MCP process or you are debugging a manual config. The agent starts `memorix serve` and communicates with it over stdio.
 
-A bare `memorix serve` defaults to `--mode micro`, which exposes 9 high-signal tools so agent context is not crowded by rarely used schemas. Use `memorix serve --mode lite` for the 20-tool solo memory surface, `--mode team` for the 28-tool coordination surface, or `--mode full` for all 47 advanced and MCP Memory compatibility tools.
+A bare `memorix serve` defaults to `--mode micro`, which exposes 9 high-signal tools so agent context is not crowded by rarely used schemas. Use `memorix serve --mode lite` for the 21-tool solo memory surface, `--mode team` for the 29-tool coordination surface, or `--mode full` for all 48 advanced and MCP Memory compatibility tools.
 
 `memorix setup` writes `--mode lite` into the installed MCP config: the generated guidance teaches the extended surface (sessions, timeline, reasoning, retention), so the installed server must actually expose it. Override with `MEMORIX_MODE` if you prefer a different profile.
 
@@ -237,6 +237,61 @@ Use foreground HTTP mode for debugging or custom launch supervision:
 memorix serve-http --port 3211
 ```
 
+`memorix background start` survives the terminal but not logout or reboot. For a supervised control plane that only execs `memorix serve-http`, use an OS unit. Keep the bind address loopback-only.
+
+Linux systemd user unit (`~/.config/systemd/user/memorix-http.service`):
+
+```ini
+[Unit]
+Description=Memorix HTTP MCP control plane
+After=default.target
+
+[Service]
+Type=simple
+ExecStart=/usr/bin/env memorix serve-http --host 127.0.0.1 --port 3211
+Environment=PATH=%h/.local/bin:/usr/local/bin:/usr/bin
+Restart=on-failure
+RestartSec=3
+
+[Install]
+WantedBy=default.target
+```
+
+```bash
+systemctl --user daemon-reload
+systemctl --user enable --now memorix-http.service
+```
+
+macOS LaunchAgent (`~/Library/LaunchAgents/com.memorix.http.plist`). Replace the `memorix` path with the output of `which memorix`:
+
+```xml
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+  <key>Label</key>
+  <string>com.memorix.http</string>
+  <key>ProgramArguments</key>
+  <array>
+    <string>/usr/local/bin/memorix</string>
+    <string>serve-http</string>
+    <string>--host</string>
+    <string>127.0.0.1</string>
+    <string>--port</string>
+    <string>3211</string>
+  </array>
+  <key>KeepAlive</key>
+  <true/>
+  <key>RunAtLoad</key>
+  <true/>
+</dict>
+</plist>
+```
+
+```bash
+launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/com.memorix.http.plist
+```
+
 Generic HTTP MCP config:
 
 ```json

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "memorix",
   "mcpName": "io.github.AVIDS2/memorix",
-  "version": "1.9.4",
+  "version": "1.9.6",
   "description": "Local-first shared memory layer for AI coding agents across MCP clients, Git history, reasoning context, and project sessions.",
   "type": "module",
   "sideEffects": false,
@@ -76,6 +76,7 @@
     "benchmark:codegraph": "node scripts/benchmark-codegraph-provider.cjs",
     "benchmark:codegraph:scale": "node scripts/benchmark-codegraph-scale.mjs",
     "benchmark:retrieval": "npm run build && node scripts/benchmark-retrieval.mjs",
+    "gate:memory-quality": "vitest run tests/evaluation/memory-quality-gate.test.ts tests/evaluation/session-replay-evaluation.test.ts",
     "gate:large-store": "npm run build && node scripts/large-store-gate.mjs",
     "sync:mcp-registry": "node scripts/sync-mcp-registry-manifest.mjs",
     "check:mcp-registry": "node scripts/check-mcp-registry-manifest.mjs",
@@ -83,7 +84,7 @@
     "check:plugin-release": "node scripts/sync-plugin-release-manifests.mjs --check",
     "audit:pi-upstream": "node scripts/pi-upstream-audit.mjs",
     "lint": "tsc --noEmit",
-    "prepublishOnly": "npm run sync:mcp-registry && npm run sync:plugin-release && npm run check:mcp-registry && npm run check:plugin-release && npm run build && npm test"
+    "prepublishOnly": "npm run check:mcp-registry && npm run check:plugin-release && npm run build && npm test"
   },
   "keywords": [
     "mcp",

package/plugins/antigravity/memorix/plugin.json

@@ -1,4 +1,4 @@
 {
   "name": "memorix",
-  "version": "1.9.4"
+  "version": "1.9.6"
 }

package/plugins/claude/memorix/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://anthropic.com/claude-code/plugin.schema.json",
   "name": "memorix",
-  "version": "1.9.4",
+  "version": "1.9.6",
   "description": "Shared workspace memory for Claude Code and other AI coding agents.",
   "author": {
     "name": "AVIDS2",

package/plugins/codebuddy/memorix/.codebuddy-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "memorix",
-  "version": "1.9.4",
+  "version": "1.9.6",
   "description": "Local-first project memory for CodeBuddy Code and other coding agents.",
   "author": {
     "name": "AVIDS2",

package/plugins/codex/memorix/.codex-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "memorix",
-  "version": "1.9.4",
+  "version": "1.9.6",
   "description": "Shared workspace memory for Codex and other AI coding agents.",
   "author": {
     "name": "AVIDS2",

package/plugins/copilot/memorix/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "memorix",
-  "version": "1.9.4",
+  "version": "1.9.6",
   "description": "Shared local workspace memory for GitHub Copilot CLI and other AI coding agents.",
   "author": {
     "name": "AVIDS2",

package/plugins/gemini/memorix/gemini-extension.json

@@ -1,6 +1,6 @@
 {
   "name": "memorix",
-  "version": "1.9.4",
+  "version": "1.9.6",
   "mcpServers": {
     "memorix": {
       "command": "memorix",

package/plugins/hermes/memorix/plugin.yaml

@@ -1,5 +1,5 @@
 name: memorix
-version: "1.9.4"
+version: "1.9.6"
 description: Shared workspace memory bridge for Hermes Agent.
 author: AVIDS2
 license: Apache-2.0