TMCRA 1.0.0-rc.1 — 一键本地记忆与可视化工作台

一键本地安装（Windows x64 预览）

下载附件 tmcra-local-windows-1.0.0-rc.1.zip，解压后双击 Install-Local.cmd。默认选择轻量档，自动准备独立 Python 环境、校验并下载模型、创建本机身份和启动记忆服务。无需 TMCRA 服务器、云端账号或预装 Python。

首次安装需要联网下载环境和模型；运行时记忆处理、模型调用与数据存储使用本机。安装后重启 TMCRA 插件/MCP 宿主，即可发现本地身份。高级 TMCRA_CONFIG_FILE 覆盖需先清除。用户仍需审核宿主插件与 Hook 权限。

档位	Embedding / Reranker	规划建议
轻量 lite-cpu（默认）	multilingual-e5-small / mmarco-mMiniLMv2-L12-H384-v1	16GB 内存，启动前约 6.3GB 空闲内存
均衡 balanced-bge	BGE-M3 / BGE-reranker-v2-m3 + TMCRA 融合	32GB 内存，检索建议 6GB+ 显存
高配候选 quality-qwen	Qwen3-Embedding-4B / Qwen3-Reranker-0.6B	64GB 内存，检索建议 16GB+ 显存

所有档位的生成默认使用本机 Qwen3-4B-Q4_K_M / llama.cpp，CPU、32K 上下文、单槽；CUDA 用于可用设备上的检索。容量是规划估算，均衡和高配尚待相应硬件实测。可运行 Install.ps1 -Profile balanced-bge 或 quality-qwen 选择档位。

本次主要更新

• 统一大版本号：npm/NuGet/插件 1.0.0-rc.1，Python 1.0.0rc1。各组件源码已对齐；未在本次发布中提供全部桌面/手机二进制包。
• 可视化记忆工作台：Writer 与后台整理模型配置、知识库与图谱查看、任务接续、会话读写模式与召回预算。
• 聊天纠错展示准确来源并等待用户确认；取消、拒绝或宿主不支持确认时保持记忆不变。
• 本地安装先登记选择，失败时阻止沿用旧云端连接；原有云端凭据与记忆保留。真实后端按校验清单复制至独立运行目录。
• 记忆服务保留完整 Source / Fast / Slow / 索引 / 召回生产流程，增加 Windows 跨进程锁和不同模型索引身份检查。

验证与限制

主项目 Release gate 全部通过。隔离目录自动创建 Python 3.12.14 环境、182 个后端文件校验、解压包安装准备流程与三个客户端本地身份发现已验证。

发布后追加核验：Codex 后台退出时的唤醒竞态已通过 PR #2 修复到 main；修复后的 13 项检查 全部通过。已发布 tag 中的 Codex 源码快照未重写；独立 Windows 后端 ZIP、DSH 与通用 MCP 不受此项影响。独立 Codex 插件仍待安全扫描问题解决后发布。

完整离线系统仍为预览：真实 CPU 单条合成记忆写入 112.218 秒、原始证据召回 0.516 秒；复杂证据编译在 600 秒超时。后台整理、完整服务重启恢复、最新内存限流参数以及均衡/高配硬件验收仍待完成。详细结果见 本地部署说明。

本次未改动生产服务器、真实用户记忆或本机已安装插件。若宿主 Agent 仍使用云端主模型，召回内容可能发送给其模型提供方；整套 Agent 离线还需要本地宿主模型。

平台发布状态

• 通用 MCP 1.0.0-rc.1 已发布，MCP Registry 同步成功。
• DSH 1.0.0-rc.1 使用相同版本线。
• 独立 Codex 插件及其市场更新暂留 PR #4：官方扫描器把 PyTorch 模型 .eval() 误判为动态执行代码，原有高风险检查门槛保留。主项目中的对应源码已更新，独立插件发布尚未完成。

Runtime ZIP SHA-256: 8f7db924db9504af01bd8ffa71c1cf63a1e3fe0836eb4ca79716e1388bfc7457.

English summary

Unified major release candidate with account-free Windows x64 local installation, bundled verified backend, automatic private Python, three pinned embedding/reranking profiles, a visual memory workspace, local provider configuration and explicitly confirmed corrections. Initial downloads need internet; local memory execution does not require TMCRA's servers. Complete offline-model acceptance is partial, as detailed above; this is a prerelease.

TMCRA v0.3.0-rc2 — Portable Production Deployment

Single-RTX-5090 clean-room validation passed with the production Qwen3.6-35B-A3B profile, full source provenance, memory graph, cross-Agent SDK sharing, and personal knowledge base.

Clean-room validation report

0.3.0-rc2 — 2026-08-20

English

• Validated the public deployment chain in an isolated single-RTX-5090 clean
  room with the production defaults: Qwen3.6-35B-A3B at three 65,536-token
  slots, four Writer workers, and two recall replicas.
• Verified ingest, read-your-writes recall, evidence provenance, memory graph,
  cross-Agent SDK sharing, and the local-Qwen personal knowledge base. The
  tested knowledge base retained a source fingerprint and seven evidence links.
• Pinned a Transformers-compatible Hugging Face Hub range and made model
  downloads portable by fetching runtime-required BGE files, disabling Xet by
  default, and checking the main BGE artifacts by SHA-256.
• Added the required BGE reranker model manifest to the standalone service
  archive and corrected the installed integrated-repository path.
• Fixed custom-prefix and custom-virtualenv resolution in service and
  maintenance controls, plus direct execution of the preflight script.
• Fixed projection-progress serialization for the valid
  dedicated-local-slot resource mode and added regression coverage.
• Kept the immutable v0.3.0-rc1 tag unchanged; these deployment fixes begin
  with v0.3.0-rc2.

中文

• 在单张 RTX 5090 隔离区按公开生产默认配置完成部署验证：
  Qwen3.6-35B-A3B 使用三个 65,536 Token 槽位，配置四个 Writer worker 与
  两个召回副本。
• 验证写入、读己之写召回、证据溯源、记忆图谱、跨 Agent SDK 共享，以及由
  本地 Qwen 生成的个人知识库；测试知识库保留来源指纹和七条证据链接。
• 固定与 Transformers 兼容的 Hugging Face Hub 版本范围；BGE 仅下载运行所需
  文件，默认禁用 Xet，并对主要 BGE 模型文件执行 SHA-256 校验。
• 将运行时要求的 BGE reranker 模型清单加入独立服务发布包，并修正安装后的
  集成源码路径。
• 修复自定义安装前缀、虚拟环境在服务控制和维护脚本中的解析顺序，同时支持
  从任意工作目录直接执行预检脚本。
• 修复投影进度接口对合法 dedicated-local-slot 资源模式的序列化，并补充
  回归测试。
• 保持不可变的 v0.3.0-rc1 标签不变；以上部署修复从 v0.3.0-rc2 起发布。

TMCRA v0.3.0-rc1 — Full-Stack Open-Source Memory Infrastructure for AI Agents

TMCRA v0.3.0-rc1 is the first public release candidate of the complete source distribution.

All TMCRA algorithms, commercial application modules, deployment automation and optimization, API/SDK/MCP adapters, prompts, knowledge graph, personal knowledge base, one-click integration, cross-Agent flows, and benchmark tooling are included in the repository.

TMCRA v0.3.0-rc1 是全量源码发行版的首个公开候选版本。

仓库包含 TMCRA 算法、商业应用模块、部署自动化与优化、API/SDK/MCP 适配器、Prompt、知识图谱、个人知识库、一键接入、跨 Agent 流程和 Benchmark 工具链。

• Release commit: 094f665643ce3d28a35a05ac07f4a05bbf222150
• Release Gate: https://github.com/reshuibuduo/tmcra/actions/runs/32346715260
• Default local production model: Qwen3.6-35B-A3B
• Reference development hardware: one RTX 5090; at least 32 GB VRAM is recommended
• Third-party weights and datasets are downloaded separately under upstream licenses

---

Changelog / 变更记录

All notable release changes are recorded here. Package ecosystems use their own
valid prerelease spelling: 0.3.0-rc.1 for npm/NuGet and 0.3.0rc1 for Python.
Both map to the repository release 0.3.0-rc1.

这里记录公开版本的重要变化。各包生态使用自身合法的预发布格式：npm/NuGet
使用 0.3.0-rc.1，Python 使用 0.3.0rc1，均对应仓库版本
0.3.0-rc1。

0.3.0-rc1 — 2026-08-20

English

• Published the complete ten-component source tree: memory algorithms, service
  API, Web console, desktop and mobile clients, SDKs and integrations, Codex and
  Claude plugins, MCP server, benchmark harness, and model/data tooling.
• Opened the production memory chain, commercial application modules, deployment
  automation, configuration templates, prompts, evidence provenance, knowledge
  graph, personal knowledge base, one-click integration, and cross-Agent flows.
• Documented the default self-hosted Qwen3.6-35B-A3B route and operator-selected
  model substitution. Exact model-name allowlists were removed; alternative
  models can require prompt and runtime tuning.
• Added source-traceable memory contracts and the adapters required to connect an
  existing memory system through the HTTP API, SDK lifecycle, MCP, or Agent hooks.
• Published the frozen LongMemEval and LoCoMo score records with model, judge,
  denominator, and single-run boundaries. The benchmark memory chain is tied to
  the 2026-04-24 DeepSeek-V4 Preview model snapshot.
• Added repository-wide version drift and credential checks plus GitHub release
  gates for API, benchmark, SDK, Web, desktop, mobile, and Android builds.
• Development reference: one NVIDIA RTX 5090. At least 32 GB VRAM is recommended
  for the default local profile. Multi-GPU topology and tuning remain deployment-
  specific and require operator validation.
• Third-party datasets and model weights remain under their upstream licenses and
  are downloaded separately; generated benchmark responses are not vendored.

中文

• 完整发布十个组件的源码：记忆算法、服务 API、Web 控制台、桌面端、移动端、
  SDK 与集成、Codex/Claude 插件、MCP Server、Benchmark 链路、模型与数据工具。
• 开源生产记忆链路、商业应用模块、部署自动化、配置模板、Prompt、证据溯源、
  知识图谱、个人知识库、一键接入和跨 Agent 流程。
• 明确默认本地生产模型为 Qwen3.6-35B-A3B，并允许部署者替换为自己的模型。
  精确型号白名单已移除；替换模型后可能需要调整 Prompt 和运行参数。
• 提供可追溯来源的记忆合同，以及通过 HTTP API、SDK 生命周期、MCP、Agent Hook
  接入现有记忆系统所需的适配层。
• 发布 LongMemEval 与 LoCoMo 的冻结成绩记录，写明模型、Judge、分母和单次运行
  边界；Benchmark 记忆链路对应 2026-04-24 DeepSeek-V4 Preview 模型快照。
• 增加全仓版本漂移检查、密钥扫描和 GitHub 发布门禁，覆盖 API、Benchmark、
  SDK、Web、桌面端、移动端与 Android 构建。
• 开发参考硬件为单张 NVIDIA RTX 5090；默认本地配置建议至少 32 GB 显存。
  多卡拓扑与调优取决于部署环境，需要部署者自行验证。
• 第三方数据集与模型权重遵循上游许可并独立下载，仓库不附带生成式评测响应。

TMCRA v0.1.1 — documentation and benchmark release

Documentation, benchmarks, and deployment release

This patch release turns the public source tree into an operator-facing guide.

Verified benchmark scorecards

• LongMemEval-S500: 411/500 (82.2%) from one frozen end-to-end 500-question run.
• Frozen Source24 production-candidate path: 77/100; Semantic V4 shadow path: 61/100 and not promoted over the baseline.
• LoCoMo: 80.92% Mem0-style LLM Judge (Categories 1–4, N=1,540, five-run mean), 55.20 official Token F1 (all 1,986), and 82.00% evidence recall (all 1,986). Protocols are reported separately.

New operator documentation

• Verified feature matrix for all 10 public components.
• Application-surface, API/runtime, commercial-module, and existing-memory-system integration/extension guides.
• Chinese engineering guide.
• Single NVIDIA RTX 5090 reference development profile, 32 GB minimum VRAM recommendation, and explicit multi-GPU operator boundary.

Integrity

• Public release audit: P0=0, P1=0, P2=0.
• 10 numbered component packages plus full-source archive.
• Full archive SHA-256: 59e02d07752edb08ca3616806bf63e2460e4763da418be27ca2a4a439251cf19.

TMCRA v0.1.0

TMCRA v0.1.0

First public source release of the self-hosted temporal memory platform for products and AI agents.

• 10 independently deployable components
• English and Simplified Chinese documentation
• Bring-your-own-memory-system integration paths
• Release archive and SHA-256 manifest included

Commit: d3cd8c6
Archive SHA-256: 419d663b7efd33982b6e917a5e91f5d1b4552f4e8ee348c8bddb13eeb0abf35f