MemoryBear v0.3.4 Community Release Notes

[keeees]

MemoryBear v0.3.4 Community Release Notes — Intelligent Connectivity

Release Date: May 20, 2026 | Codename: LingTong (灵通 · Intelligent Connectivity)

MemoryBear v0.3.4 Community weaves identity, intelligence, and integration into a single coherent fabric. This release introduces a full SSO + LingQi commerce loop for tenant onboarding and subscription management, deepens memory intelligence with temporal awareness and graph-relation retrieval, opens up the ontology engineering layer to external systems, and brings end-to-end transparency to Agent execution across all node types. Together these changes make MemoryBear meaningfully easier to operate at scale while improving the everyday experience of building and reasoning over memory.

---

🚀 I. Core Upgrade Overview

1. SSO & LingQi Integration 🔌

• Package Viewing API: SSO now exposes an endpoint for LingQi to query available subscription packages, allowing identity-side dashboards to surface plan information without round-tripping through the product UI.
• Tenant Creation with Default Free Plan: The SSO tenant creation interface automatically provisions a free plan, syncs the corresponding order record, and links the tenant to its primary contact, removing several manual steps from onboarding.
• OIDC API Key Provisioning: After a workspace is created via SSO, OIDC offers a dedicated interface to mint that workspace's API key, so machine-to-machine credentials are established as part of the same identity flow.
• Package Purchase API with Source Tagging: SSO provides an endpoint for LingQi to place subscription orders on behalf of a tenant, with a source marker on each order so finance and analytics can attribute revenue back to the originating channel.

---

2. Memory Intelligence 🧠

• Language-Aware Image Input: Image inputs are now processed in the user's input language, so multimodal turns return descriptions and extracted facts that match the conversation's working language rather than defaulting to English.
• Reflection Engine Description Merging: The reflection engine consolidates duplicate or near-duplicate description fields when summarizing memories, producing tighter reflections and reducing noise in downstream retrieval.
• Timestamped Entity Descriptions: Entity description fields now carry timestamps, making it possible to reason about how an entity's understanding evolved and to surface the most recent characterization during recall.
• Entity-Relation Retrieval: Memory search can now traverse entity relationships, enabling queries like "what did the user discuss about Project X with their manager" to be answered through graph hops rather than pure vector similarity.
• Accurate Text-to-File Entity Linking: Entity nodes extracted from typed user messages are now correctly associated with entities discovered in uploaded files, ensuring a single coherent graph instead of parallel disconnected ones.
• Metadata Field Refresh: The memory metadata schema has been updated with refined fields and validation, improving filtering precision and downstream search relevance.
• Ontology External Service API: Ontology Engineering exposes a public service API, letting external systems read and operate on the ontology layer that previously sat behind internal-only interfaces.

---

3. Application & Workflow ⚙️

• Pointer Mode One-Click Duplicate Selection: When a node needs to be replicated across an application, pointer mode now selects all instances of the same node in one click, so bulk edits no longer require hunting through the canvas.
• Voice & Reaction Interactions: Conversations now support voice playback of replies, like/dislike feedback, and follow-up question generation that proposes natural next prompts based on the current turn.
• Agent Execution Visualization: The frontend now streams Agent execution steps as they happen, turning long waits from a black box into a visible progression and giving end users confidence that the system is making progress.
• Improved Web Search: The web search node has been retuned for better source selection, freshness, and ranking, producing more relevant snippets for grounding.
• Per-Node Agent Logs: Agent logs are now scoped per node, so each step of an execution can be inspected independently — invaluable for debugging long pipelines.
• Document Extractor with Optional Image Parsing & URL Preservation: The document extractor now offers a configurable toggle for image parsing and preserves URLs in extracted text, so links survive the extraction step intact.
• Per-Node Data Processing Trace with Single-Node Run: Every node type — HTTP, LLM, parameter extraction, question classification, knowledge base, code execution, tools, conditional branch, variable assignment, memory extraction, memory write, and document extractor — now exposes its inputs, outputs, and intermediate processing, with single-node run for fast iteration.
• Custom Application Tags: Applications can be created with user-defined tags that double as filters in the application list, making large workspaces much easier to navigate.
• Tool System Accuracy Tuning: The tool-calling pipeline has been tuned for higher fidelity in tool selection and argument generation, reducing missed or malformed calls during model invocations.

---

4. Knowledge Base Integration 📚

• Multimodal Image Parsing — Chinese Output: Multimodal image parsing produces noticeably more natural Chinese descriptions, fixing cases where Chinese output read as machine-translated.
• One-Click Knowledge Base Export: Knowledge bases support multi-select and one-click batch download of source files, simplifying backup, migration, and audit workflows.
• Cleaner XLSX QA Imports: When importing XLSX in QA mode, sheet names are no longer appended to each QA pair, leaving the imported content clean and faithful to the source.

---

5. Robustness & Bug Fixes 🔧

• Memory UI Loading Performance: Resolved slow loading in the memory dashboard caused by oversized initial fetches; the view now paginates and lazy-loads, restoring snappy interaction on large graphs.
• Alias Extraction Display: Fixed a distillation-stage bug where the user's extracted alias was overwritten with the literal string "user"; aliases are now preserved as-extracted.
• Visual-Capability Mismatch in Serial Model Nodes: Fixed an error where two LLM nodes connected in series would fail when the upstream node had vision enabled and the downstream node did not; the runtime now reconciles modalities at the node boundary.
• Deep Thinking Toggle on Hybrid Models: Disabling deep thinking on hybrid thinking models previously kept the model thinking silently — slow without visible benefit. Hybrid models now bypass the thinking phase entirely for direct responses, while pure-thinking models simply suppress the thinking output. Configurable in custom model settings and per-application model settings.
• PDF Parsing via mineru: Fixed PDF parsing failures on the mineru engine, restoring reliable ingestion for documents that were previously rejected.

---

🧭 Looking Ahead

MemoryBear v0.3.4 Community is the release where the platform stops being just a memory engine and becomes a system you can operate. SSO with LingQi closes the loop from identity to subscription to provisioning, OIDC-managed API keys remove a class of credential-handling friction, and tenant onboarding becomes a single transaction rather than a checklist. For teams running their own deployments, this is the version where MemoryBear fits into the existing identity and commerce stack instead of standing beside it.

The intelligence layer takes a quieter but equally important step. Timestamped entity descriptions and relationship-based retrieval move memory beyond similarity search into something closer to reasoning over a living knowledge graph. Combined with language-aware multimodal input and consistent text-to-file entity linking, recall feels less like keyword matching and more like a system that actually remembers context. The new ontology external service API opens this layer to integrators, inviting other systems to build on top of MemoryBear's structural understanding rather than replicating it.

The third theme is transparency. Per-node logs, single-node runs, and live Agent execution visualization turn application building from guesswork into engineering. Builders can see what each step did, replay individual nodes in isolation, and watch the trace as it streams. Combined with tool-system accuracy improvements and the deep-thinking toggle fix, applications behave the way their authors expect — and when they don't, the answer to "why" is one click away.

The knowledge base, meanwhile, gets the kind of quiet polish that operations teams notice immediately: more natural multimodal Chinese output, one-click batch export for backups and audits, cleaner XLSX QA imports, and a restored mineru PDF path for documents that previously got stuck at the door.

The next version will deepen this trajectory: richer ontology APIs, more dimensions of temporal and causal reasoning in memory recall, deeper visual debugging for branched workflows, and more knowledge-base ingestion formats with stronger multimodal support. Expect ongoing performance work across the memory and application stacks, and continued investment in the open API surface for the community.

---

---

MemoryBear v0.3.4 社区版 发布说明 —— 智慧通达

发布日期： 2026年5月20日 ｜ 版本代号： 灵通（LingTong · Intelligent Connectivity）

MemoryBear v0.3.4 社区版将身份、智能与集成织成一张完整的网。本版本引入了 SSO 与灵启的完整商业闭环，覆盖租户开通到套餐订购的全链路；为记忆智能加入时间感知与关系图检索，将本体工程层开放给外部系统；并在所有节点类型上实现 Agent 执行过程的端到端可视。这些变化让 MemoryBear 更适合规模化运营，也让日常的应用搭建与记忆推理体验更加顺畅。

---

🚀 一、核心升级概览

1. SSO 与灵启集成 🔌

• 套餐查询接口：SSO 新增供灵启调用的套餐查询接口，身份侧仪表板无需绕回产品端即可获取套餐信息。
• 租户创建与默认免费套餐：SSO 租户创建接口在新建租户时自动分配免费套餐、同步相应订单并关联主联系人，省去开通过程中多个手动环节。
• OIDC API Key 创建：通过 SSO 创建空间后，OIDC 提供专用接口为该空间签发 API Key，机器间凭证在身份流程中一并完成。
• 套餐购买接口（含来源标记）：SSO 提供供灵启代租户下单购买套餐的接口，订单上携带来源标记，便于财务与运营追踪订阅来源。

---

2. 记忆智能 🧠

• 图片输入语言对齐：图片输入按用户当前输入语言处理，多模态轮次返回的描述与抽取结果与会话语言一致，不再默认英文。
• 反思引擎描述合并：反思阶段对重复或近似的 description 字段进行合并，反思结果更紧凑，下游召回噪声更低。
• 实体描述时间戳：实体的 description 字段加入时间戳，可观察实体认知的演变，并在召回时优先展示最新描述。
• 基于实体关系的检索：记忆检索支持沿实体关系跳转，"用户与上级讨论过 Project X 的内容"这类问题能通过图谱跳跃回答，而非仅靠向量相似。
• 文字与文件实体精准对齐：用户输入文本中抽取的实体节点与上传文件中识别的实体正确合并到同一图谱，不再形成两套互相隔离的关系。
• metadata 字段更新：记忆 metadata 字段进行结构与校验更新，过滤更精准，搜索相关性进一步提升。
• 本体工程对外服务 API：本体工程对外开放服务接口，第三方系统可直接读取并操作此前仅限内部使用的本体层。

---

3. 应用与工作流 ⚙️

• 指针模式一键选中相同节点：复制相同节点时，指针模式可一键选中画布中所有同类节点，批量编辑不再需要逐个查找。
• 语音播报与互动反馈：对话支持语音播报、点赞 / 踩反馈，以及基于当前回答自动生成的追问建议。
• Agent 执行过程可视化：前端实时流式展示 Agent 的执行步骤，长等待从黑盒变为可见进度，让终端用户对系统状态心中有数。
• 联网搜索优化：联网搜索节点对来源筛选、新鲜度与排序进行调优，引证素材更具相关性。
• 节点级 Agent 日志：Agent 日志按节点维度组织，可逐节点查看输入输出，长流程的排障效率显著提升。
• 文档提取器图片解析开关与 URL 保留：文档提取器新增图片解析开关，并在提取过程中保留原文 URL，链接不会在中间环节丢失。
• 节点数据处理过程与单节点运行：HTTP、LLM、参数提取、问题分类、知识库、代码执行、工具、条件分支、变量赋值、记忆提取、记忆写入、文档提取器等全部节点均可查看其输入输出与中间处理，并支持单节点运行用于快速调试。
• 自定义应用标签：创建应用时可设置自定义标签，标签同时作为应用列表的筛选条件，便于在大规模工作区中查找。
• 工具系统调优：工具调用流程经过调优，工具选择与参数生成准确率提升，模型调用中的漏调与错调显著减少。

---

4. 知识库集成 📚

• 多模态图片解析中文输出优化：多模态图片解析的中文输出更自然，纠正了此前部分中文结果"机翻感"明显的问题。
• 一键导出知识库：知识库支持多选与源文件批量下载，备份、迁移与审计流程更便捷。
• XLSX QA 导入更干净：QA 模式导入 XLSX 时不再在每条 QA 后追加 sheet 名称，导入内容忠实于源文件。

---

5. 稳健性与缺陷修复 🔧

• 记忆界面加载性能：修复因初始数据集过大导致的记忆面板加载缓慢问题，改为分页与按需加载，大型图谱下交互响应恢复流畅。
• 用户别名提取显示：修复萃取阶段用户别名被错误覆盖为字面"用户"的问题，别名按抽取结果原样保留。
• 模型节点串行视觉能力错配：修复两个 LLM 节点串联时，上游开启视觉而下游未开启所导致的报错，运行时在节点边界自动协调模态。
• 混合思考模型深度思考开关：此前关闭深度思考时，混合思考模型仍在隐式思考——响应慢而无可见收益。混合思考模型现在可完全跳过思考阶段直接响应，纯思考模型则仅抑制思考内容输出。可在自定义模型配置与应用模型配置中启用。
• mineru 引擎 PDF 解析：修复 mineru 解析引擎下 PDF 解析失败问题，此前被拒的文档可正常入库。

---

🧭 未来展望

MemoryBear v0.3.4 社区版是平台从"记忆引擎"转向"可被运营的系统"的关键版本。SSO 与灵启的联动打通了身份—订阅—开通的完整链路，OIDC 托管的 API Key 消除了一类凭证处理摩擦，租户开通从一份清单变成一次事务。对自部署团队而言，这是 MemoryBear 真正融入既有身份与商业体系的版本，而不再是并立其旁。

智能层的进展更安静但同样关键。实体描述的时间戳与基于关系的检索，让记忆从相似度检索走向更接近"基于活体知识图谱进行推理"的形态。结合图片输入的语言对齐与文字—文件实体的稳定合并，记忆召回不再只是关键词匹配，而是真正"记得上下文"的系统。本体对外服务 API 进一步把这一层开放给集成方，让第三方系统在 MemoryBear 的结构化理解之上构建，而非另起炉灶。

第三条主线是透明度。节点级日志、单节点运行与 Agent 执行的实时可视化，让应用搭建从猜测变成工程：每一步都可见，每一节点都可重放。配合工具系统准确性提升与深度思考开关修复，应用的行为更贴近作者的预期，遇到问题也能一步定位原因。

知识库则收获了运维团队会立刻察觉的那种安静的打磨：更自然的多模态中文输出、面向备份与审计的一键批量导出、更干净的 XLSX QA 导入，以及恢复正常的 mineru PDF 解析路径——此前被拒的文档不再卡在门口。

下一个版本将沿着这条线继续推进：更丰富的本体 API、记忆召回中更多维度的时间与因果推理、对分支工作流更深入的可视化调试，以及覆盖更多格式与多模态能力的知识库摄入。我们会持续投入记忆与应用栈的性能优化，并持续扩充开放给社区的 API 表面。

---

This discussion was created from the release MemoryBear v0.3.4 Community Release Notes.