loveRhythm1990 · Oct 28, 2026 · Oct 28, 2026 · Oct 28, 2026
diff --git a/README.md b/README.md
@@ -3,12 +3,12 @@
 [![CI - Lint, Tests & Adapter Build](https://github.com/matrixorigin/memory_benchmarks/actions/workflows/aml-ci.yml/badge.svg?branch=main)](https://github.com/matrixorigin/memory_benchmarks/actions/workflows/aml-ci.yml)
 
 Public memory benchmarks, an Agent Memory Leaderboard (AML) Add/Search adapter,
-and a dedicated dev deployment for [Memoria](https://github.com/matrixorigin/memoria).
+and deployment templates for [Memoria](https://github.com/matrixorigin/memoria).
 
 This repository helps the team evaluate how well Memoria retrieves evidence from
 conversation histories. It prepares public datasets, writes the original history
 through `/add`, queries `/search`, and records retrieval metrics and run checkpoints.
-It also owns the adapter and its Aliyun dev deployment pipeline.
+Deployment templates are included; live environments have their own release process.
 
 **Default mode: retrieval diagnostics.** LoCoMo and LongMemEval-S also support
 optional local Answer/Eval. Native evaluator parity, coding scores and official
@@ -22,8 +22,9 @@ AML leaderboard scores remain separate acceptance work. See the
 | Try the deployed API | Adapter access key | [API quick start](#try-the-deployed-api) |
 | Run a benchmark in Actions | Workflow access; configured adapter access key | [Run a benchmark](#run-a-benchmark) |
 | Develop the adapter locally | Python 3.12+, uv, running Memoria, master key | [Local development](#local-development) |
+| Reproduce a fixed deployment | Pinned mb/Memoria versions and configured backend | [Deployment and reproduction](deploy/README.md#本地复现) |
 | Deploy or update services | Configured deployment Secrets and Variables | [Deployment](#deploy-or-update-services) |
-| Understand the design | No runtime credentials | [Competition design](agent_memory_leaderboard/docs/competition-plan.md) |
+| Understand the design | No runtime credentials | [Architecture](#how-it-works) and [adapter guide](agent_memory_leaderboard/README.md) |
 
 ## How it works
 
@@ -115,9 +116,8 @@ not automatically clean it up.
 | `GET /docs` | Interactive request/response documentation |
 
 Healthy probes alone do not validate authentication, model calls, or retrieval
-quality. The dev deployment and seven-check live API integration smoke passed on
-2026-10-06; the [acceptance record](agent_memory_leaderboard/docs/competition-plan.md#当前交付状态)
-contains the actual revisions and remaining validation scope.
+quality. Verify these paths against the intended environment before a benchmark;
+this guide does not certify a particular deployment or official evaluation run.
 
 For the complete field contract and Python example, see
 [public integration](deploy/README.md#接入已部署的-addsearch). For retry and error
@@ -170,8 +170,9 @@ not consume official AML Full quota.
 
 See the [benchmark workflow guide](agent_memory_leaderboard/benchmarks/CI.md)
 for data provenance, metric definitions, timeout and resume behavior.
-Recorded self-run results and their limitations are available in
-[baseline reports](agent_memory_leaderboard/benchmarks/baselines/README.md).
+Keep environment-specific baseline reports, checkpoints and run notes under the
+Git-ignored `var/` directory. Record the code, dataset and configuration used; do
+not treat local results as official AML scores.
 
 ### Interpreting the metrics
 
@@ -245,17 +246,14 @@ local integration smoke, benchmark commands, and a Memoria Docker Compose overla
 
 ## Deploy or update services
 
-QA services are managed in **moi-gitops**, stack `idc`: the `aml-qa` unit owns the
-adapter and the `memoria` unit owns shared Memoria. See the
-[QA deployment handbook](https://github.com/matrixorigin/moi-gitops/blob/main/docs/handbooks/qa-aml-adapter.md).
-The old Aliyun direct-deployment workflow has been removed. This repository's
-active CI/benchmark workflows do not deploy to Aliyun or require its cluster,
-SQL or Memoria administration credentials. Legacy deployment code and its
-[reference guide](deploy/README.md) remain for historical reference.
+See the [deployment reference](deploy/README.md) for the template layout,
+configuration boundaries and Add/Search integration. The active workflows in
+this repository run checks and optional QA benchmarks; they do not deploy an
+application. Shared environments use their existing reviewed release process.
 
-The adapter image build → moi-gitops image-update PR workflow is follow-up work.
-ACR still hosts the application images, and the configured qwen-plus answer/judge
-service still uses DashScope. These dependencies are separate from hosting QA.
+Deployment plans, environment-specific acceptance records and internal run notes
+are local operator material, rather than public usage documentation. Keep them
+under Git-ignored `var/private-docs/`; do not link public guides to local files.
 
 ## Workflows at a glance
 
@@ -268,13 +266,13 @@ service still uses DashScope. These dependencies are separate from hosting QA.
 
 ```text
 memory_benchmarks/
-├── .github/workflows/          # Offline CI, dev deployment, manual benchmark
+├── .github/workflows/          # Offline CI and manual benchmark
 ├── agent_memory_leaderboard/
 │   ├── src/                   # Adapter, journal, HTTP client, shared benchmark program
 │   ├── benchmarks/            # Parsers, preparation, entrypoints and CI guide
 │   ├── scripts/               # Local launchers and Actions preparation/reporting
 │   ├── tests/                 # Adapter and benchmark tests using simulated backends
-│   ├── docs/                  # Maintained design and review references
+│   ├── docs/                  # Public feature and experiment guides
 │   └── compose.adapter.yaml   # Overlay for Memoria Docker Compose
 └── deploy/
     ├── chart/                 # Helm resources for Memoria and adapter
@@ -289,9 +287,9 @@ when obtaining or sharing its data.
 
 ## Design and contribution
 
-The maintained [competition design](agent_memory_leaderboard/docs/competition-plan.md)
-records architecture, decisions, verified delivery status, and remaining work.
-Detailed guides own operational instructions; this README is the project entrypoint.
+The architecture overview above and the public adapter, deployment and benchmark
+guides describe supported behavior. Keep internal plans, live acceptance records
+and experiment outputs in Git-ignored local storage.
 
 Explain the design, validation evidence, and effect on API behavior or evaluation
 comparability in each PR. Preserve provenance and user isolation, keep gold data

diff --git a/agent_memory_leaderboard/.gitignore b/agent_memory_leaderboard/.gitignore
@@ -8,3 +8,17 @@ var/
 benchmarks/*/*/data/
 benchmarks/*/*/results/
 state/*.json
+
+# Local retry journals and their SQLite sidecars.
+*.db
+*.db-journal
+*.db-wal
+*.db-shm
+*.sqlite
+*.sqlite-journal
+*.sqlite-wal
+*.sqlite-shm
+*.sqlite3
+*.sqlite3-journal
+*.sqlite3-wal
+*.sqlite3-shm
diff --git a/agent_memory_leaderboard/README.md b/agent_memory_leaderboard/README.md
@@ -59,25 +59,13 @@ It checks adapter liveness, not Memoria/database/embedding readiness.
 `GET /ready` checks journal readability and the upstream Memoria health endpoint,
 without writing memories or invoking a model. It is a dependency-availability
 check, not evidence that Add/Search or the embedding provider works correctly.
-The deployment target is the Aliyun dev unit cluster, using a dedicated Memoria
-backend and database login. Local evaluation remains available.
-See [competition design](docs/competition-plan.md) for runtime scope and workload
-budgets. The offline CI checks code and builds the adapter image without
-publishing or deploying it. Memoria source builds run only in the dev deployment
-workflow when updating Memoria in source mode, including preview. Adapter-only
-updates and existing-image mode skip Memoria builds.
-The [dev deployment workflow](../deploy/README.md) defaults to building Memoria
-from its `main` branch; another branch, tag, commit, or existing image can be selected.
-It publishes the selected images and deploys both services by digest through an
-independent Pulumi stack, recording the actual source commit for each source build.
-Public access requires DNS; the existing cert-manager ClusterIssuer automatically
-issues and renews the TLS Secret.
-
-The Aliyun dev endpoint is deployed at `https://aml.cn-dev.matrixone.tech`.
-See the [deployment steps and Actions inputs](../deploy/README.md#发布流程) and
-[public curl/Python integration examples](../deploy/README.md#接入已部署的-addsearch).
-Public clients need only the adapter API key; they do not need Memoria master,
-database, or Kubernetes credentials.
+For local development, configure a running Memoria service and use the adapter's
+loopback endpoint. Hosted environments supply their own endpoint, capacity and
+release configuration; see the [deployment reference](../deploy/README.md).
+The offline CI checks code and builds the adapter without deploying it. The
+manual QA benchmark calls its configured QA service and does not select or
+update a deployment. Public clients use only the adapter key; database,
+Kubernetes and Memoria administration credentials stay with the operators.
 
 ## Contract verification
 
@@ -90,9 +78,9 @@ For submission, use public endpoint URLs without embedded credentials; the
 localhost examples below are for development. Enable authentication for Full.
 Configure proxy/load-balancer request timeouts for the documented 30-minute
 request window. `MEMORIA_TIMEOUT_SECONDS` controls each upstream call (default
-120 seconds), not the total Add duration. The deployed dev endpoint has passed
-the project's live integration smoke; official platform Smoke and long-request
-validation remain pending. See the [current acceptance status](docs/competition-plan.md#当前交付状态).
+120 seconds), not the total Add duration. A local smoke, a deployment health
+check and an official AML Smoke are separate checks; keep their run records and
+version bindings outside the public usage guide.
 
 Concurrent Add/Search requests can queue native Memoria calls without changing
 synchronous Add semantics. Set `MEMORIA_MAX_CONCURRENCY=1` and

diff --git a/agent_memory_leaderboard/benchmarks/CI.md b/agent_memory_leaderboard/benchmarks/CI.md
@@ -61,14 +61,14 @@ benchmark 请求 QA 当前部署的 Add/Search；Summary 与 JSON 记录环境
 旧阿里云部署 workflow 已移除。QA 使用 `aml-qa-benchmark`
 只串行本仓库的 QA 评测；moi-gitops 在另一个仓库，同名 concurrency group 也不能跨仓库互斥。
 QA 评测前需要协调共享 Memoria 与 GitOps 发布；本 workflow 不宣称已自动冻结它们。
-QA gateway 初始限制一个活跃业务请求、每秒一个后端业务请求，完整历史可能明显慢于 dev。
+实际入口和后端限制以目标环境当前配置为准；运行前核对容量，不能从历史配置推断本轮能力。
 并发组不阻止其他本地客户端、AML 官方任务或手工操作；GitHub 默认只保留一个 pending
 运行，新 pending 可能替换旧 pending，不应同时排队多个部署／评测任务。
 
 ## Summary 与完整报告
 
-已完成运行的安全聚合指标与解释见[固定基线](baselines/README.md)。完整断点仍在
-对应 Actions artifact；Git 中的聚合报告不包含数据正文、回答或可恢复的 private cache。
+具体运行的基线、版本和验收记录保存在 Git 忽略的本地 `var/` 目录或受控运行记录中。
+公开文档仅说明如何生成和解释报告，不绑定某次环境验收结果。
 
 Summary 包含状态、评测进程耗时、写入分块数、已处理／已查询／跳过的问题数、空结果、
 有证据标签的问题数、召回率、Search P50／P95、按数据集 category 分组的结果及版本。
@@ -144,24 +144,17 @@ Summary 单列 eligible／scored／pending、评分覆盖率及两类失败题
 缺失时不写入记忆。准备步骤不读取模型凭据。工作流安装锁定的 evaluation extra；
 适配器镜像不需要此 extra，也不需要重新部署。
 
-当前开发配置（2026-10-07）：回答和裁判的 MODEL 均为 `qwen-plus`，两个 BASE_URL
-均为 `https://dashscope.aliyuncs.com/compatible-mode/v1`，分别配置上述两个 API_KEY
-Secrets，复用已有百炼凭据。GitHub 配置已保存并回读确认；尚未验证本次凭据的模型
-调用或执行新契约基准。密钥不进入仓库或日志。
+根据所选模型服务填写上述 Variables 和 Secrets。示例配置不代表某个仓库已经配置凭据，
+也不代表模型调用或完整评测已通过验证。密钥不进入仓库或日志。
 
-运行 `AML benchmark` 时选择 `main`、`dataset=locomo`、`mode=answer-eval`、
+运行 `QA memory benchmark` 时选择 `main`、`dataset=locomo`、`mode=answer-eval`、
 `histories=1`、`top_k=100`，`runtime_label` 可留空。模型设置在 benchmark 进程中生效，
 不需要重新构建或部署 Memoria／适配器。qwen-plus 为提供商可更新别名，跨日期比较
 不能假定底层权重相同；长期可比实验需选择并记录可用快照版本。回答与裁判同模型的
 分数用于本地迭代，不能作为独立裁判或官方评分等价性的证明。
 
-正式 AML 的 Answer/Eval 由平台固定，我们的配置不会替换平台模型。官方 Full checklist
-写 Open-source Methods 的 Add 期望使用 `gpt-4o-mini`，商业产品无 Add/Search 模型限制；
-FAQ 仍写学术榜 embedding 为 `text-embedding-v4`、LLM 组件为 `gpt-4o-mini`、reranker
-不限制。两套组别名称的对应及纯存储 Add／上海 bge-m3 资格需向官方确认；不能用本地
-qwen-plus 配置推断正式参评合规。
-依据：[AML 文档](https://agentmemories.ai/zh-cn/docs)、
-[赛事 FAQ](https://agentmemoryleaderboard.ai/competition/)。
+正式 AML 的回答与评分由平台提供，本地配置不会替换平台模型。参赛组别和模型要求
+以[官方文档](https://agentmemoryleaderboard.ai/docs)为准，本地运行结果不证明正式参评合规。
 
 全部输入（指令、角色、问题、选项／问题日期、证据）按固定 `o200k_base` 编码加
 256 token 余量进行本地预算计数，上限 117760；保留完整记录的排名前缀。它不是所有
@@ -210,7 +203,7 @@ Answer/Eval 的 private cache 也受同一保留期约束，不能随报告公
 报告按现有保留策略记录清理期限；120 分钟步骤被强制终止、runner 未执行 finally 时，
 最终报告步骤补写未设置的保留期限，不延长已有期限。
 Artifact 过期不清理 Memoria、用户库、journal 或历史快照；物理清理需要独立运维方案。
-正式评测前执行发布冻结及保留管理流程，见[比赛方案](../docs/competition-plan.md)。
+正式评测前与环境维护者协调发布冻结、版本记录和保留管理，见[部署参考](../../deploy/README.md)。
 
 该 workflow 消耗我们自己的 embedding、数据库和 Actions 资源，不消耗 AML 官方 Full
 配额。请求串行发送；限制 histories 和执行时长不等于完整的全局 QPS／token 预算。
@@ -224,10 +217,9 @@ Artifact 过期不清理 Memoria、用户库、journal 或历史快照；物理
 | 作答（仅 answer-eval） | `BENCHMARK_ANSWER_API_KEY` | `AML_BENCHMARK_ANSWER_API_KEY` |
 | 判分（仅 answer-eval） | `BENCHMARK_JUDGE_API_KEY` | `AML_BENCHMARK_JUDGE_API_KEY` |
 
-本地完整 QA 配置保存在私有 `var/evaluation-qa.env`（当前独立 worktree 下），权限 0600、
-Git 忽略；模型名和 base URL 也已配置。GitHub 的上述三个 Secret 已配置，模型变量沿用
-qwen-plus / DashScope。本地真实 QA 最小链路已通过；GitHub-hosted runner 的端到端执行
-需在本 PR 合入后首次运行验证，Secret 已存在不等于运行已验证。
+本地配置可保存在 Git 忽略的 `var/evaluation-qa.env`，使用仅当前用户可读写的权限；
+模型名称、地址和凭据由操作者自行配置。不要将凭据是否存在写成链路验证结论，
+也不要把本地 private cache 或内部验收记录加入仓库。
 
 离线 CI 不使用任何业务 key。评测不需要 KUBECONFIG、Pulumi token、数据库密码、
 Memoria master key 或 embedding key；这些由 QA 后端和 GitOps 持有。当前服务运行于

diff --git a/agent_memory_leaderboard/benchmarks/EVALUATION.md b/agent_memory_leaderboard/benchmarks/EVALUATION.md
@@ -41,7 +41,7 @@ flow, but the complete official judge prompts and a reproducible implementation
 were not found in the documentation checked on 2026-10-07. The prompts in
 `src/judges/` come from the public benchmark repositories, not from AML.
 See the [official documentation](https://agentmemories.ai/zh-cn/docs) and
-[competition readiness plan](../docs/competition-plan.md#正式参评前的工作).
+[deployment reference](../../deploy/README.md).
 
 ## LongMemEval answer reader