MemoryBear v0.3.5 Community Release Notes — Tempered Edge

MemoryBear v0.3.5 Community Release Notes — Tempered Edge

Release Date: May 27, 2026 | Codename: MoJian (磨剑 · Sharpening the Sword)

MemoryBear v0.3.5 Community is a sharpening release. Building on the broad surface introduced in v0.3.4, this version refines memory recall and writing into a faster, more nuanced flow, expands node-level capabilities across Agents and workflows, brings real multimodal interaction into embedded scenarios, and tightens the Open API surface for SSO and external integrations. A long list of bug fixes — including several with cross-space and import scenarios — round out the release.

---

🚀 I. Core Upgrade Overview

1. Memory Intelligence 🧠

• Retrieval & Response Speed: Memory quick-retrieval and the normal response path have been optimized end-to-end, reducing latency for both warm recall and standard turns so conversations feel noticeably more responsive.
• Adaptive Memory Writing: The memory write pipeline now supports adjustable write position, sliding writes, delayed writes, and long-context reference resolution, giving the engine room to capture meaning rather than just dumping turns into storage.
• Memory Pruning Engine Update: Pruning now covers AI messages, long user text segments, and reference resolution for user-uploaded images, keeping the memory layer focused on signal rather than verbatim history.
• Reflection & Pruning On by Default: New workspaces have the reflection engine and pruning configuration enabled out of the box, so the memory layer behaves intelligently from the first conversation rather than waiting for opt-in.
• Reflection Entity Deduplication: The reflection engine now deduplicates and disambiguates entities during summarization, preventing the gradual accumulation of near-duplicate nodes in long-running workspaces.
• Predicate & Entity Type Enum + Fuzzy Output: Entity relation edge predicate values and entity_type values now support both numeric enumerations and fuzzy type outputs, making both strict and exploratory schemas first-class.
• Async Memory Write API for API Keys: API-key holders gain an asynchronous memory write endpoint, decoupling ingestion from request latency for high-throughput integrations.
• LocOMO Test Set Validation: Continued optimization against the LocOMO benchmark with verified improvements, anchoring memory quality to a reproducible reference rather than ad-hoc test cases.

---

2. Application & Workflow ⚙️

• LLM Node — Parameter Completion, Retry, Inference Separation: The LLM node gains automatic parameter completion, configurable failure retry, and a clear separation between the inference (thinking) phase and the response phase, making long-running calls both more reliable and easier to debug.
• Parameter Extractor Advanced Inference Mode: The parameter extractor exposes an advanced setting that lets users pick between Function Calling and Prompt-based instruction modes based on how each model responds best, improving extraction fidelity across model families.
• Question Classifier with Vision: The question classifier node now supports vision-capable models, so multimodal turns can be routed correctly without falling back to a text-only proxy.
• Variable Node Append for Array[Object]: Variable nodes gain an operation type. When a variable is created as Array[Object], clicking the variable node now exposes an append action, making structured accumulation across a workflow straightforward.
• Agent Variable Shortcut Reference in Prompts: Variables defined under an Agent can now be referenced directly in prompts via a shortcut, removing the previous detour of routing them through intermediate nodes.
• Application Annotation on Shared Applications: Applications (Agent, workflow) now carry an annotation field, and shared applications display the annotation correctly — fixing the prior gap where shared instances surfaced without their notes.
• Model Marketplace & Model Combination Layout Swap: The relative positions of the Model Marketplace and Model Combination panels have been swapped to reflect actual usage frequency, putting the more common entry point in front.
• Application List Filter by Type: The application list supports filtering by application type, so large workspaces can be navigated by Agent / workflow / other categorical lenses rather than scrolled.

---

3. Multimodal & Interaction 💬

• iframe Multimodal Embedding: Embedded iframe deployments now support full multimodal interaction including text, image, and voice — restoring the parity with the legacy mid-platform that some integration scenarios depended on.
• Conversation Sharing & Feedback Toolkit: Shared conversations support per-message share links, selection-based feedback and reporting, message deletion, and regeneration with a random seed to produce alternative outputs from the same prompt.
• Multimodal Files Visible in Logs: Logs now display user-sent multimodal attachments such as images, closing a gap where investigators could see the message but not the actual file the user uploaded.
• Mobile-Adaptive Share Links: MemoryBear share links now adapt to mobile viewports rather than rendering as a desktop layout shrunk to fit, restoring legibility on phones.

---

4. Knowledge Base 📚

• QA Edit-Then-Download: Knowledge base entries in QA format can be edited in-place and downloaded with the edits applied, so review workflows produce a clean export without needing to track changes externally.
• CSV Dataset Query Reliability: Knowledge bases containing CSV datasets now reliably surface in retrieval — fixing prior cases where CSV-backed entries were difficult to recall.
• Retrieval Parameter Refactor: Keyword retrieval now uses similarity_threshold and semantic retrieval uses vector_similarity_weight, replacing the previous overlapping naming and making each parameter's role unambiguous.

---

5. Open API 🔌

• /v1 vs /api Memory Write Body Consistency [1007376]: The body JSON format for the /v1 external memory write endpoint has been aligned with the internal /api endpoint — particularly the message field — removing a class of integration mismatches.

---

6. Robustness & Bug Fixes 🔧

• [1007667] Cross-Space Knowledge Base Recall via Agent: Fixed a bug where a knowledge base shared from space A to space B could be recalled in B's recall test but produced no chunks when bound to a newly created Agent in B. Agent-side retrieval now resolves shared KBs the same way the test interface does.
• [1007725] Application Config Import Consistency: Fixed mismatches when importing application configs — Agent imports could lose their memory enabled state, and workflow imports with mismatched memory-node configurations imported silently. Both paths now preserve state, and import-time mismatches surface a prompt.
• [1007683] Shared Agent Knowledge Base Tool Name: Fixed an issue where Agents shared across spaces displayed an incorrect knowledge-base tool name, breaking tool selection in the recipient space.
• [1007476] KB Recall Test — Similarity Zero Should Not Recall: Fixed an edge case where the KB recall test surfaced results with similarity score 0. The threshold is now strictly applied so zero-similarity matches are excluded.
• [1006882] Multi-Space Member Chrome Login: Tracked fix for an unreproduced report of multi-space members being unable to load the login page in Chrome; defensive handling has been added on the affected code path.
• Dify Workflow Import — Hyphenated IDs: Fixed a workflow import path where IDs imported from Dify containing hyphens (e.g., code-split) were rendered as a code variable reference. Hyphen-bearing identifiers are now preserved literally.
• Log Content Search Speed: Log search by content has been optimized for large log volumes, reducing wait times when investigating an issue across long conversation histories.

---

🧭 Looking Ahead

MemoryBear v0.3.5 Community is the version where the platform stops chasing surface and starts pressing edges. v0.3.4 finished assembling the operational picture; v0.3.5 sharpens nearly every part of it. Memory writes are smarter about position and timing, retrieval is faster and more discriminating, the Agent and workflow node families pick up the kind of small affordances — variable shortcuts, append actions, parameter completion, retry, vision-aware routing — that turn a builder's mental model into actual mechanics. None of these alone is a headline; together they change how the platform feels under daily use.

The forward-looking significance is most visible in two places. First, the multimodal interaction surface: bringing voice, image, and text into iframe embeds, surfacing user-sent files in logs, and adapting share links to mobile viewports together signal that MemoryBear is being treated as a system that lives across devices and embedded contexts, not just a desktop application. Second, the Open API surface: aligning /v1 and /api body formats and introducing async memory writes for API keys together describe a platform that is becoming a dependable backbone for upstream integrators rather than an island.

The next version will continue along this line: deeper temporal and causal reasoning over the entity graph, broader vision support across application nodes, more configurability around memory pruning and reflection cadence, and continued tightening of the Open API surface — alongside ongoing performance work in retrieval and a steady cadence of community-reported bug fixes.

---

---

MemoryBear v0.3.5 社区版 发布说明 —— 百炼成锋

发布日期： 2026年5月27日 ｜ 版本代号： 磨剑（MoJian · Sharpening the Sword）

MemoryBear v0.3.5 社区版是一次"打磨型"发布。在 v0.3.4 拓宽平台表面的基础上，本版本将记忆召回与写入打磨成更快、更精细的流水线，扩展 Agent 与工作流节点的能力边界，将真正的多模态交互带入嵌入式场景，并在 SSO 与对外集成方向...

MemoryBear v0.3.4 Community Release Notes

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...

MemoryBear v0.3.3 Community Release Note

MemoryBear v0.3.3 Community Release Notes — Expanding Realms

Release Date: May 14, 2026 | Codename: YanJing (衍境 · Expanding Realms)

MemoryBear v0.3.3 Community delivers a comprehensive set of upgrades to memory intelligence, knowledge base management, and workflow orchestration. This release introduces temporal retrieval for memory entities, a revamped extraction pipeline, full single-node workflow execution, and significant knowledge base enhancements — alongside a broad set of stability fixes across the platform.

---

🚀 I. Core Upgrade Overview

1. Memory Intelligence 🧠

• Dashboard End-User Filtering: The dashboard/end_users endpoint now filters out users with zero memory entries, keeping the dashboard focused on active memory profiles.
• New Extraction Pipeline: Switched to the new memory extraction pipeline, delivering improved accuracy and throughput for entity and relationship extraction.
• Ontology Upper/Lower Level Update: Updated the universal ontology engineering definitions for upper and lower levels, resulting in more accurate entity and relationship category extraction.
• Metadata Display Enhancement: The /memory-storage/end_user_info endpoint now surfaces only the most relevant user profile fields: goals, traits, interests, and core_facts.
• Alias Extraction Improvement: Aliases are now extracted during node merging when alias relationships exist, ensuring alias information stays current as the knowledge graph evolves.
• Temporal Fields for Entity Relations: Entity relationships now carry three time fields — dialog_at (conversation time), valid_at (event occurrence time), and invalid_at (event expiration time) — enabling precise temporal reasoning.
• L0 Memory Read Enhancement: Memory reads now return fixed user alias and user description fields (goals, traits, interests, core_facts) as part of the L0 response.
• Temporal Retrieval (valid_at): Statements can now be retrieved by valid_at event occurrence time, enabling time-based memory queries such as "what happened last week."
• Memory Validation Session Support: The memory validation endpoint now accepts session_id and removes the history parameter, enabling proper coreference resolution across conversation turns.
• Knowledge Base Custom Prompt for QA Extraction: Mind-map knowledge base QA pair extraction now supports custom prompt input — uses system prompt when empty, custom prompt when provided.

---

2. Knowledge Base Integration 📚

• CSV Tabular Import Logic: CSV file imports now follow a one-QA-pair-per-chunk model, where each row of question-answer data generates an independent chunk for precise retrieval.
• Manual QA Chunk Fix: Fixed the issue where manually creating a knowledge entry with QA extraction would generate duplicate entries; added chunk deletion support (with graph rebuild trigger button).
• Visible IDs on Knowledge Page: Knowledge base pages now display knowledge base ID, document ID, and chunk ID for easier debugging and cross-referencing.
• File System Abstraction: Knowledge base file storage now uses the API-level file system, supporting OSS, S3, and local storage backends interchangeably.
• Batch Operations API: Added server-side batch operation APIs for knowledge base entry management, enabling bulk imports and modifications.

---

3. Workflow & Application ⚙️

• Single-Node Execution Phase 1: Workflows now support running individual nodes in isolation — covering Model and Tool node types for rapid testing and debugging.
• Single-Node Execution Phase 2: Extended single-node execution to all remaining node types: Code Execution, HTTP Request, Memory Read, Knowledge Retrieval, Memory Storage, Parameter Extraction, Document Extractor, List Operations, Template Rendering, Question Classifier, Iteration, Loop, Variable Aggregation, Variable Assignment, and Conditional Branch.
• Variable Real-Value Display: Node inputs containing variables now display their resolved real values instead of variable names in the node visualization panel.
• Model Node Exception Branch: Model nodes now include an exception handling branch to gracefully handle sensitive word filtering errors instead of failing silently.
• OpenAI Channel Tool Parameter Fix: Fixed the Agent OpenAI channel tool parameter passing issue caused by model gateway provider misconfiguration (provider incorrectly set to openai for gateway-sourced models).

---

4. Frontend & UX 🎨

• Safari Workflow Support: Workflow editor now fully supports Safari browser, resolving rendering and interaction issues specific to WebKit.
• Login Page UI Redesign: Completely redesigned the login page with a modern, streamlined interface.

---

5. Robustness & Bug Fixes 🔧

• Multi-Space Member Login (Chrome): Fixed an issue where members belonging to multiple spaces could not access the login page on Chrome.
• Workflow Share Node Error Interaction: Improved error handling for the workflow experience-sharing node to display proper error feedback instead of failing silently. (In Progress)
• Hybrid Retrieval Threshold Mismatch: Resolved inconsistency between the knowledge base recall test's similarity thresholds and the workflow knowledge retrieval node's vector similarity thresholds.
• Vector Similarity Weight Drift: Fixed an issue where vector_similarity_weight would change unexpectedly during speech library testing with the knowledge retrieval node in 400-inbound call workflows.
• Reranker Model Config Failure: Fixed the knowledge retrieval node's global configuration where the Reranker model selector would fail to open.
• JSON Output Toggle Reset: Fixed a bug where selecting a JSON-output model, enabling the JSON toggle, clicking away, and clicking back would reset the toggle — and saving would not persist the setting.
• Search Mode Parameter Leakage: Fixed an issue where switching from hybrid search to keyword or semantic search would retain both similarity_threshold and vector_similarity_weight parameters, even though only one applies to the selected mode.
• Memory Read Validation Error: Fixed the deep-thinking mode memory extraction error where asking "who am I" caused a Pydantic validation failure (string_type expected but received list input ['用户是谁']).
• HTTP Request Streaming Error: Fixed the HTTP request node exception "Attempted to access streaming request content, without having called read()" by properly handling streaming response bodies.

---

🧭 Looking Ahead

MemoryBear v0.3.3 represents a significant step toward production-grade memory intelligence. The introduction of temporal fields, enhanced extraction pipelines, and L0 memory enrichment collectively transform MemoryBear from a memory store into a temporally-aware reasoning substrate — one that understands not just what happened, but when it happened and when it stopped being true.

The workflow single-node execution capability marks a turning point for developer experience. By enabling isolated testing of every node type, we dramatically reduce the iteration cycle for complex workflow development. Combined with real-value variable display, this release makes MemoryBear workflows significantly more debuggable and portable.

The next version will focus on a comprehensive upgrade of the memory engine's core capabilities, completing application interaction and Agent/workflow features, optimizing knowledge base and frontend iframe multimodal support, introducing SSO/OIDC integration interfaces, and resolving remaining known issues across the platform.

---

---

MemoryBear v0.3.3 社区版 发布说明 —— 衍境

发布日期： 2026年5月14日 ｜ 版本代号： 衍境（YanJing · Expanding Realms）

MemoryBear v0.3.3 社区版对记忆智能、知识库管理和工作流编排进行了全面升级。本版本引入了记忆实体的时间检索能力、全新萃取流水线、完整的工作流单节点执行支持以及多项知识库增强，同时修复了平台中大量稳定性问题。

---

🚀 一、核心升级概览

1. 记忆智能 🧠

• Dashboard 用户过滤：dashboard/end_users 接口现在会过滤记忆数为 0 的用户，保持仪表盘聚焦于活跃记忆档案。
• 切换新萃取流水线：切换至新一代记忆萃取流水线，提升实体和关系提取的准确率与吞吐量。
• 本体工程 Upper/Lower Level 更新：更新通用本体工程的上下层级定义，现在能提取到更准确的实体/关系类别。
• Metadata 内容展示增强：/memory-storage/end_user_info 接口现在只展示最相关的用户画像字段：goals、traits、interests、core_facts。
• 别名提取改进：现在会在有别名关系的节点合并时自动提取别名，确保知识图谱演进过程中别名信息保持最新。
• 实体关系时间字段：实体关系新增三个时间字段——dialog_at（对话发生时间）、valid_at（事件发生时间）、invalid_at（事件失效时间），支持精确的时间推理。
• L0 记忆读取增强：记忆读取时现在会返回固定的用户别名和用户描述字段（goals、traits、interests、core_facts）。
• 时间检索（valid_at）：现在可以通过 valid_at 事件发生时间检索 statement 句子，支持"上周发生了什么"等时间维度查询。
• 记忆验证 Session 支持：记忆验证接口新增 session_id 参数，去除 history 参数，以支持跨对话轮次的代词消融。
• 知识库自定义提示词 QA 提取：脑图知识库问答对提取支持自定义提示词输入——未填写时采用系统提示词，填写则采用自定义提示词。

---

2. 知识库集成 📚

• CSV 表格导入逻辑：CSV 文件导入后按一条 QA 对应一个分块（chunk）的方式处理，即一行问答数据生成一个独立分块，确保检索精度。
• 手动 QA 分块修复：修复手动创建知识选择问答对提取会生成两条的问题；支持删除分块（删除后需重建图谱，增加了操作按钮）。
• 知识库页面 ID 展示：知识库页面现在显示知识库 ID、文档 ID 和 chunk ID，方便调试和交叉引用。
• 文件系统抽象：知识库文件存储使用 API 层文件系统，支持 OSS、S3 和本地存储后端互换。
• 批量操作 API：新增知识库录入的服务端批量操作 API，支持批量导入和修改。

---

3. 工作流与应用 ⚙️

• 单节点运行（第一阶段）：工作流支持单独运行模型节点和工具节点，实现快速测试和调试。
• 单节点运行（第二阶段）：扩展单节点执行至所有剩余节点类型：代码执行、HTTP 请求、记忆提取、知识检索、记忆存储、参数提取、文档提取器、列表操作、模版渲染、问题分类器、迭代、循环、变量聚合、变量赋值、条件分支。
• 变量真实值展示：节点输入中包含变量时，现在展示解析后的真实值而非变量名称。
• 模型节点异常处理分支：模型节点新增异常处理分支，优雅处理敏感词过滤等错误，避免静默失败。
• OpenAI 通道工具传参修复：修复 Agent OpenAI 通道工具传参问题，原因是模型网关来源的模型 provider 被错误设置为 openai。

---

4. 前端与用户体验 🎨

• Safari 工作流支持：工作流编辑器全面支持 Safari 浏览器，解决 WebKit 特有的渲染和交互问题。
• 登录页 UI 改版：全新设计的登录页面，界面更加现代简洁。

---

5. 稳定性与缺陷修复 🔧

• 多空间成员登录（Chrome）：修复属于多个空间的成员无法在 Chrome 上访问登录页面的问题。
• 工作流分享节点报错交互：改进工作流体验分享节点的错误处理，显示正确的错误反馈。（开发中）
• 混合检索阈值不匹配：解决知识库召回测试中相似度阈值与工作流知识检索节点中向量相似度阈值不一致的问题。
• 向量相似度权重漂移：修复 400 呼入话术库测试时知识检索节点 vector_similarity_weight 意外变化的问题。
• Reranker 模型配置失败：修复知识检索节点全局配置中 Reranker 模型选择器无法打开的问题。...

MemoryBear v0.3.2 Community Release Notes — Sharpening the Blade

MemoryBear v0.3.2 Release Notes — Sharpening the Blade

Release Date: April 29, 2026 | Codename: LingFeng (凌锋 · Edge of the Summit Wind)

MemoryBear v0.3.2 builds upon the foundation of v0.3.1 with a sweeping set of improvements across memory intelligence, application workflows, and frontend stability. This release sharpens the platform's core capabilities — from paginated explicit memory queries and optimized retrieval paths, to deep workflow observability and robust error handling. With over 30 items spanning new features, UX refinements, and critical bug fixes, v0.3.2 represents a significant step toward production-grade resilience and developer ergonomics.

---

🚀 I. Core Upgrade Overview

1. Memory Intelligence 🧠

• Explicit Memory Pagination & Query Filters: The explicit memory API now supports full pagination along with query conditions including time range and important event filtering, enabling efficient browsing and targeted retrieval of large memory stores.
• Memory Store Detail API (Batch): A new batch-capable memory store detail API allows retrieving detailed information for multiple memory stores in a single request, reducing round-trips and improving dashboard responsiveness.
• Memory Retrieval Path Optimization: The memory retrieval pipeline has been restructured and optimized to deliver more relevant and accurate results. This change touches the core search path and requires thorough testing to validate improvements across diverse query patterns.
• User-Level Fair Scheduling: Memory processing tasks are now distributed using a user-level fair scheduling algorithm, ensuring equitable resource allocation across tenants and preventing any single user from monopolizing background processing capacity.

---

2. Workflow & Application Engine ⚙️

• Agent Context Coherence Fix: Resolved a critical issue where the Agent would exhibit context confusion — reading prior messages incorrectly and generating incoherent responses. Conversation state management has been tightened to maintain strict turn-by-turn coherence.
• Workflow Trial Run Node Flow Visualization: During workflow trial runs, the UI now displays real-time node transition flow, making it easy to observe which nodes are executing and in what order.
• Per-Node Workflow Logging: Application logs now include granular per-node log entries for workflow executions, showing the flow of data through each node in the pipeline.
• Log Content Search & Filtering: Application log viewing now supports full-text content search and filtering, allowing developers to quickly locate specific events or errors within large log volumes.
• HTTP Request Node cURL Display: HTTP request nodes in workflows now display the exact cURL command being sent, providing full transparency into outbound API calls for debugging and auditing.
• Third-Party Workflow Import Variable Adaptation: When importing third-party workflows, HTTP request node body JSON now correctly adapts session variables, ensuring seamless integration without manual variable remapping.
• URL-Based Model Input Support: Models are no longer restricted to sys.files for document input — direct URL passing is now supported, enabling reading of document content from URLs. Supports docx and pdf formats. The workflow document extractor can identify image positions within documents and process them into file objects with placeholder formatting like [图片 第2页 第1张]: http://..., with an added images file object output.
• Workflow Import with END Node Output: Workflow import now fully supports the workflow type including END node output configuration, enabling complete round-trip import/export of complex workflow definitions.
• Knowledge Base Citation Download: When conversations involve knowledge base retrieval, referenced documents can now be downloaded directly. This feature can be toggled on/off per application.
• Workflow Undo/Redo Support: Added undo and redo functionality for workflow editing, supporting Ctrl+Z keyboard shortcuts and providing UI buttons for reverting accidental changes.
• Skill Keyword Made Optional: The keyword field for skills has been changed from required to optional, reducing friction when creating skills that don't need keyword-based triggering.
• Tool Schema Body Parameter Support: Tool creation schemas now properly support Body parameters in addition to Params, resolving input length limitations encountered in scenarios like markdown-to-word conversion.

---

3. Frontend & User Experience 🎨

• Streaming Output Abort Support: Frontend streaming output now supports request termination, allowing users to stop long-running LLM responses mid-stream.
• Prompt Generation Flicker Fix: Resolved the flickering issue during Agent prompt generation and shared conversations caused by frequent page redraws during LLM streaming output. The rendering pipeline has been stabilized to eliminate visual jitter.
• Model Management Tab State Fix: Fixed an issue where rapidly switching tabs in the model management page would cause the active tab indicator to disappear or display incorrectly.
• Configuration Checklist Reactivity Fix: Clearing configuration items now immediately triggers checklist validation updates, rather than requiring a manual refresh.
• Model List Tag Display Fix: Fixed truncated tag display in the model list view, ensuring all tags are fully visible.

---

4. Security, Permissions & API 🔒

• API Key Creation Gate: Applications that have not been published can no longer have API keys created for them, preventing premature API access to draft applications.
• Deleted Knowledge Base Citation Cleanup: When a knowledge base bound to an agent or workflow is deleted, published versions no longer display stale "citation and attribution" references in experience sharing and API call contexts.
• API Key Creation with Rate Limits: API key creation now includes QPS and daily call limit configuration, enforcing plan-level caps (up to 50 for the current tier).
• Super Admin Workspace Admin Creation Fix: Fixed an error where creating a workspace administrator as a super admin would result in incorrect tool, skill, and model associations.
• Invited User Tenant ID Fix: When inviting a user to a workspace who then registers, their tenant_id is now correctly set to match the inviter's tenant, resolving cascading errors with tool, skill, and model access that resulted from mismatched tenant assignments.

---

5. Model & Runtime Resilience 🔧

• Model Group Limit Enforcement Fix: After a model group exceeds its combination limit, individual models in the list can now be properly enabled again, resolving a state lock that previously prevented model activation.
• Deep Thinking Token Overflow Handling: When Agent deep thinking mode produces context that exceeds the maximum token limit (observed with doubao-seed-2-0-mini-260215), the system now handles the overflow gracefully instead of returning a truncated response with no body content.

---

🧭 Looking Ahead

MemoryBear v0.3.2 marks a pivotal moment in the platform's journey toward production-grade maturity. The breadth of this release — spanning memory retrieval optimization, workflow observability, fair scheduling, and deep bug fixes — reflects a system that is being battle-tested in real-world deployments and refined based on concrete operational feedback. The focus on resilience, from token overflow handling to tenant isolation fixes, signals a platform that takes reliability as seriously as capability.

The introduction of user-level fair scheduling and optimized memory retrieval paths lays the groundwork for MemoryBear to scale gracefully under multi-tenant production loads. Combined with the new workflow node-level logging and cURL transparency, developers now have unprecedented visibility into what the platform is doing and why — a critical requirement for enterprise adoption and debugging complex AI agent behaviors.

Looking forward, the next releases will deepen workflow orchestration capabilities, expand memory intelligence with more sophisticated retrieval and ranking algorithms, and continue hardening the platform's multi-tenant architecture. Expect further improvements in real-time collaboration features, advanced permission models, and tighter integration with external tool ecosystems.

---

MemoryBear v0.3.2 发布说明 —— 锋从磨砺出

发布日期： 2026年4月29日 ｜ 版本代号： 凌锋（LingFeng · Edge of the Summit Wind）

MemoryBear v0.3.2 在 v0.3.1 的基础上进行了全面升级，涵盖记忆智能、应用工作流和前端稳定性等多个维度。本版本磨砺平台核心能力——从显性记忆分页查询与检索路径优化，到深度工作流可观测性与健壮的错误处理。超过30项功能新增、体验优化和关键缺陷修复，v0.3.2 标志着平台向生产级稳健性和开发者友好性迈出了坚实一步。

---

🚀 一、核心升级概览

1. 记忆智能 🧠

• 显性记忆分页与查询条件：显性记忆接口现已支持完整分页功能，并新增时间范围、重要事件等查询条件，支持在大规模记忆库中高效浏览和精准检索。
• 记忆库详情批量API：新增批量记忆库详情API，支持单次请求获取多个记忆库的详细信息，减少网络往返，提升仪表盘响应速度。
• 记忆检索路径优化：对记忆检索管线进行了结构性重构和优化，以提供更相关、更准确的检索结果。此变更涉及核心搜索路径，需充分测试以验证在多样化查询场景下的改进效果。
• 用户级公平调度：记忆处理任务现采用用户级公平调度算法进行分发，确保租户间资源公平分配，防止单一用户独占后台处理能力。

---

2. 工作流与应用引擎 ⚙️

• Agent上下文紊乱修复：修复了Agent上下文混乱的严重问题——此前Agent会错误读取历史消息并生成不连贯的回复。会话状态管理已加强，确保严格的逐轮对话一致性。
• 工作流试运行节点流转可视化：工作流试运行期间，界面现可实时展示节点流转过程，便于观察各节点的执行顺序和状态。
• 工作流节点级日志：应用日志现包含工作流执行的逐节点日志条目，展示数据在各节点间的流转详情。
• 日志内容搜索与过滤：应用日志查看现支持全文内容搜索和过滤，开发者可在大量日志中快速定位特定事件或错误。
• HTTP请求节点cURL展示：工作流中的HTTP请求节点现可展示实际发送的cURL命令，为调试和审计提供完整的出站API调用透明度。
• 第三方工作流导入变量适配：导入第三方工作流时，HTTP请求节点body中的JSON会话变量现可正确适配，无需手动重新映射变量即可无缝集成。
• URL直传模型输入支持：模型输入不再局限于sys.files——现支持直接传入URL读取文档内容，兼容docx和pdf格式。工作流文档提取器可识别文档中图片位置并处理为文件对象，以[图片 第2页 第1张]: http://...形式占位，并新增images文件对象输出。
• 工作流导入支持END节点输出：工作流导入现已完整支持workflow类型，包括END节点输出配置，实现复杂工作流定义的完整导入导出。
• 知识库引用文档下载：对话涉及知识库检索时，引用的文档现可直接下载。此功能可在应用级别开启或关闭。
• 工作流撤销/重做支持：新增工作流编辑的撤销和重做功能，支持Ctrl+Z快捷键操作，并提供UI按钮用于回退误操作。
• 技能关键词改为非必填：技能的关键词字段已从必填改为选填，降低创建无需关键词...

MemoryBear v0.3.1 Community Release Notes — The Boundless

MemoryBear v0.3.1 Community Release Notes — The Boundless

Release Date: April 22, 2026 | Codename: WuJing (无境 · The Boundless)

MemoryBear v0.3.1 Community Edition delivers a focused set of application-level improvements, memory API extensibility, and workflow reliability fixes. This release introduces model JSON output control, memory API key provisioning, enhanced workflow validation, and numerous UX refinements — dissolving boundaries between components and integrations to let capability flow freely.

---

🚀 I. Core Upgrade Overview

1. Application & Model Enhancements

• Model Auto-Disable on Key Deletion: When all API keys for a model are deleted, the model is now automatically disabled, preventing runtime errors from keyless model references.
• Model JSON Formatted Output Toggle: Model JSON structured output can now be toggled on and off. The official JSON control parameter ensures more stable JSON output — critical for migrating legacy middleware workflows that depend on strict JSON formatting. Custom models also support JSON output configuration.
• Configuration Import with Override: Application configurations now support import-with-override, allowing users to replace the current configuration entirely from an exported file.
• Import Cleanup for Missing Resources: When importing an application, if a referenced tool does not exist, the tool node configuration is automatically cleared. Similarly, missing knowledge bases are cleaned up, preventing broken references in imported workflows.

---

2. Memory API & Intelligence 📚

• Memory Read/Write API with End-User Key Provisioning: Memory read and write operations are now accessible via API, with support for creating end_user API keys and retrieving memory configuration API keys — enabling third-party integrations to interact with the memory layer directly.
• Memory Store API & Configuration Update: The memory store now exposes API endpoints for memory retrieval and configuration updates (with sequential interface support), providing programmatic control over memory settings.
• End-User Metadata Storage: A new storage scheme for end_user metadata has been implemented, allowing richer user context to be persisted alongside memory entries.

---

3. Workflow & UX Improvements ⚙️

• Conversation History File Metadata: Conversation history file entries now include file size, file name, and specific type labels (experience sharing, working memory), providing richer context in chat history.
• Iteration Node Parallel Input Fix: Fixed an issue where iteration nodes did not correctly handle parallel input, restoring expected concurrent execution behavior.
• API Key Last Four Digits Display: API keys now display their last four characters across the platform (including knowledge base and memory engine keys), improving key identification without exposing full secrets.
• Condition Branch Multi-File Sub-Variables: Condition branch nodes in MemoryBear now support adding sub-variables to multi-file variable types, enabling more granular conditional logic.
• Agent Model Config Reset Endpoint: Added the missing backend API endpoint for the "Reset to Default" action in Agent model configuration, completing the frontend-backend contract.
• Three-Level Variable Keyboard Navigation: Three-level variable selectors in workflows now support keyboard up/down arrow navigation, improving the variable selection experience.
• Dynamic Tab Title for Applications: Opening a new application in MemoryBear's application management now dynamically sets the browser tab title to the application name, replacing the previous static "记忆熊" text.
• Variable Aggregator Three-Level Checkbox Fix: Fixed an issue where three-level variable checkboxes in the variable aggregator node did not behave correctly.
• Workflow Checklist Validation Enhancements: The workflow checklist now includes tool node required-tool validation and LLM node vision-enabled required-variable validation, catching configuration errors before execution.
• Variable Aggregator to Parameter Extractor Output: Fixed a bug where a parameter extractor node following a variable aggregator node could not access the aggregator's output variables.

---

4. Knowledge Base & Performance ⚡

• Async Document Parsing & Graph Execution: Document upload parsing and graph processing in document_task now execute asynchronously, with increased thread counts for work-document and work-graph workers, significantly improving document ingestion throughput.

---

5. Robustness & Bug Fixes 🔧

• Tool Node Raw Parameter Types: Tool nodes now send request parameters with their original values and types instead of converting everything to strings, fixing type mismatch issues in downstream tool integrations.
• Stale Frontend Resource Import Error: Resolved an issue where, after a frontend deployment, users with cached stale resources would encounter import errors when the system attempted to load outdated data chunks.
• Workflow Tool Node Required Validation: The workflow checklist now enforces that tool nodes have their required tool configured, and LLM nodes with vision enabled must have their vision variable set — preventing incomplete configurations from being published.

---

🧭 Looking Ahead

MemoryBear v0.3.1 represents a pivotal moment in the platform's philosophical evolution — the dissolution of boundaries. Where previous releases built walls of capability, this release tears them down: SSO tenant management, subscription plan configuration, and workspace-scoped API key provisioning create a seamless fabric where identity, billing, and resource isolation coexist without friction. Every enhancement was driven by the demands of real-world enterprise adoption, yet guided by the vision that true power lies in openness.

The expanded memory API surface — with end-user key provisioning, metadata storage, and extraction pipeline tuning — signals MemoryBear's trajectory toward becoming an open, programmable memory infrastructure. Third-party systems can now interact with the memory layer as a service, not just as an embedded component. This shift from closed to open memory access is the philosophical core of this release: boundaries are illusions, and capability should flow like water.

Looking forward, we will deepen the SaaS management capabilities with usage analytics, billing integration, and tenant-level quota enforcement. The memory intelligence pipeline will continue to evolve with improved extraction accuracy, cross-memory reasoning, and adaptive forgetting strategies. Beyond boundaries, beyond limits — the boundless awaits.

---

MemoryBear v0.3.1 社区版 发布说明 —— 无境

发布日期： 2026年4月22日 ｜ 版本代号： 无境（WuJing · The Boundless）

MemoryBear v0.3.1 社区版带来了一系列应用层改进、记忆 API 可扩展性提升和工作流可靠性修复。本版本引入了模型 JSON 输出控制、记忆 API Key 供给、增强的工作流校验以及众多体验优化——打破组件与集成之间的边界，使能力自由流动。

---

🚀 一、核心升级概览

1. 应用与模型增强

• 模型 Key 全删后自动关闭：当模型的所有 API Key 被删除后，该模型将自动关闭，避免无 Key 模型引用导致的运行时错误。
• 模型 JSON 格式化输出开关：模型 JSON 结构化输出现可开关控制。官方 JSON 控制参数可确保更稳定的 JSON 输出——这对迁移依赖严格 JSON 格式的旧中台工作流至关重要，强制迁移会导致 JSON 格式报错。自定义模型同样支持 JSON 输出配置。
• 配置导入覆盖：应用配置现支持导入覆盖，用户可从导出文件完整替换当前配置。
• 导入时缺失资源清理：导入应用时，若引用的工具不存在，工具节点配置将自动清空；知识库不存在时同样清空，防止导入后出现断裂引用。

---

2. 记忆 API 与智能 📚

• 记忆读写 API 与 End-User Key 供给：记忆读写操作现可通过 API 调用，支持创建 end_user API Key 和获取记忆配置 API Key——使第三方集成可直接与记忆层交互。
• 记忆库 API 与配置更新：记忆库现开放 API 端点，支持记忆检索和配置更新（提供顺序接口），提供对记忆设置的程序化控制。
• End-User 元数据存储：实现了 end_user 元数据的存储方案，允许更丰富的用户上下文与记忆条目一同持久化。

---

3. 工作流与体验优化 ⚙️

• 会话历史文件元数据：会话历史文件条目现包含文件大小、文件名称和具体类型标签（体验分享、工作记忆），提供更丰富的聊天历史上下文。
• 迭代节点并行输入修复：修复迭代节点未正确处理并行输入的问题，恢复预期的并发执行行为。
• API Key 后四位展示：API Key 现在全平台展示后四位字符（包括知识库和记忆引擎 Key），在不暴露完整密钥的前提下便于识别。
• 条件分支多文件子变量：记忆熊中条件分支节点现支持为多文件变量类型添加子变量，实现更精细的条件逻辑。
• Agent 模型配置重置接口：补充了 Agent 模型配置"重置默认"操作缺失的后端 API 接口，完善前后端契约。
• 三级变量键盘导航：工作流中三级变量选择器现支持键盘上下箭头导航，提升变量选择体验。
• 应用标签页动态标题：在记忆熊应用管理中打开新应用时，浏览器标签页标题现动态显示应用名称，替代之前固定的"记忆熊"文本。
• 变量聚合三级勾选修复：修复变量聚合器节点中三级变量勾选行为不正确的问题。
• 工作流检查清单校验增强：工作流检查清单现包含工具节点必填工具校验和 LLM 节点视觉开启时视觉变量必填校验，在执行前捕获配置错误。
• 变量聚合器到参数提取器输出：修复参数提取器节点跟在变量聚合器节点后无法获取聚合器输出变量的问题。

---

4. 知识库与性能 ⚡

• 文档解析与 Graph 异步执行：document_task 中的文档上传解析与 Graph 处理现异步执行，同时增加 work-document 和 work-graph 的线程数，显著提升文档摄入吞吐量。

---

5. 稳健性与缺陷修复 🔧

• 工具节点原始参数类型：工具节点现以原始值及类型发送请求参数，而非全部转换为字符串，修复下游工具集成中的类型不匹配问题。
• 前端部署后资源过期导入错误：解决前端应用部署后，用户设备上缓存的过期资源在系统尝试导入旧数据块时导致导入错误的问题。
• 工作流工具节点必填校验：工作流检查清单现强制要求工具节点配置必填工具，LLM 节点开启视觉时必须设置视觉变量——防止不完整配置被发布。

---

🧭 未来展望

MemoryBear v0.3.1 是平台哲学演进中的关键时刻——边界的打破。此前的版本构筑能力之墙，而本版本将其拆解：SSO 租户管理、套餐计划配置和空间级 API Key 供给编织出一张无缝的网络，身份认证、计费与资源隔离在其中和谐共存。每一项增强都源于真实企业落地的需求，却始终以"真正的力量在于开放"为指引。

记忆 API 能力的扩展——End-User Key 供给、元数据存储和萃取流水线调优——标志着 MemoryBear 正朝着开放、可编程的记忆基础设施方向演进。第三方系统现可将记忆层作为服务调用，而非仅作为嵌入式组件。从封闭到开放的记忆访问转变，正是本版本的哲学内核：边界皆为幻象，能力当如水般自由流动。

展望未来，我们将深化 SaaS 管理能力，引入用量分析、计费集成和租户级配额管控。记忆智能管线将持续演进，提升萃取精度、跨记忆推理和自适应遗忘策略。破界而行，臻于无境。

MemoryBear v0.3.0 Enterprise Release Notes — Daybreak

Release Date: April 15, 2026 | Codename: PoXiao (破晓 · Daybreak)

MemoryBear v0.3.0 builds upon the foundation of previous releases with a broad set of improvements spanning application workflows, memory intelligence, and system robustness. This release introduces versioned API support, multimodal memory perception, and significant workflow enhancements — while resolving a wide range of stability issues across the platform. The result is a more resilient, precise, and developer-friendly MemoryBear.

---

🚀 I. Core Upgrade Overview

1. Application & API Enhancements

• Versioned API Support: External service APIs now support specifying a version when making calls, enabling consumers to pin to stable interfaces and facilitating smoother API evolution without breaking existing integrations.
• Workflow Checklist: A new workflow checklist feature provides structured validation steps, helping users verify workflow configurations before deployment and reducing runtime errors.
• Deep Thinking Parameter Control: OpenAI deep thinking parameters are now only sent to models that explicitly support deep reasoning, preventing unnecessary parameter injection into incompatible models and avoiding potential errors.
• Prompt Optimizer Model Return Optimization: Improved the response handling of the prompt optimizer model, delivering more consistent and reliable outputs during prompt engineering workflows.

---

2. Memory Intelligence 🧠

• Multimodal Memory Perception Agent: The memory perception agent now supports reading and writing multimodal memory, enabling richer context capture across text, image, and other modalities for more comprehensive user understanding.
• OpenClaw Built-in Tool: Added a new built-in tool to OpenClaw, expanding the toolkit available for memory-related agent operations and simplifying integration workflows.

---

3. User Experience 🎨

• Streaming Render Stabilization: Resolved frequent page redraws during LLM streaming output that caused visible page jitter. The rendering pipeline is now smoother, providing a stable reading experience during model responses.
• Memory Hub Renaming: The workspace section previously labeled "记忆相关" (Memory Related) has been renamed to "记忆中枢" (Memory Hub), better reflecting its role as the central memory management interface.

---

4. Workflow Improvements ⚙️

• Three-Level Variable Template Conversion: Workflow template conversion now supports three-level variables, fixing an issue where single-file-type node selections could not be properly resolved.
• VL Model Token Tracking: Token usage is now correctly tracked when using VL (Vision-Language) models from model groups, resolving a gap where tokens went unrecorded compared to selecting the same model from the standard model list.
• Imported Workflow Feature Sync: Workflows imported from external sources now correctly synchronize all feature properties including opening messages, citations, and attribution settings — previously these were silently dropped during import.
• Session Variable Name Uniqueness: Added uniqueness validation for session variable and start node variable names in workflows, preventing silent conflicts and hard-to-debug runtime issues.
• File Type Extraction Fix: Workflows can now correctly extract file.type information, resolving a parsing issue that previously returned empty values.
• Condition Branch Display Fix: Condition branch nodes now render correctly when configured with a value of 0 or when referencing session variables, fixing a display truncation issue.
• Object/Array Validation Rules: Added validation rules for object and array[object] session variables, preventing JSON serialization errors that previously blocked workflow saves.
• HTTP Request Body Key Fix: In workflow HTTP request nodes, the body field now uses key instead of name for field identification, aligning with standard HTTP conventions and resolving mapping issues.

---

5. Knowledge Base 📚

• Embedding Token Truncation Safety: Fixed errors caused by embedding model token limits being exceeded. A unified 8,000-token truncation safety boundary is now applied to all embedding models. Additionally, Excel row data now maintains independent chunks without merging, using newline separation for cleaner document processing.

---

6. Robustness & Bug Fixes 🔧

• Atomic Update & Batch Access Failures: Resolved failures in atomic memory updates and batch access record operations that could cause data inconsistency under concurrent load.
• Alias Extraction in Conversations: Fixed incorrect alias extraction during application conversations, ensuring user identities are properly resolved from dialogue context.
• Workflow Alias Extraction: Corrected alias extraction in workflows to properly distinguish between user and AI responses, including consideration of emotion and interest distribution (issue #1006016).
• RAG Memory Pagination: Fixed data pagination issues in RAG memory retrieval that could return incomplete or duplicated results across pages.
• Implicit Memory Detail Display: Resolved an issue where implicit memory entries with corresponding numeric values showed empty content in the detail view (observed in workflow v0.2.10 multimodal memory reads).
• Vector Query Driver Closed: Fixed search_graph_by_embedding: statements vector query exceptions caused by a prematurely closed database driver in memory storage and memory verification flows.
• User Management Enable/Disable: Fixed abnormal API request behavior when toggling user enable/disable status in user management.
• Model List Filter Inconsistency: Resolved an issue where model list filter criteria did not match the displayed results, ensuring accurate filtering behavior.

---

🧭 Looking Ahead

MemoryBear v0.3.0 marks a meaningful step toward production maturity. The breadth of this release — spanning API versioning, multimodal memory, workflow robustness, and rendering stability — reflects a platform that is rapidly hardening its core while continuing to expand its capabilities. The focus on fixing edge cases and improving developer-facing consistency signals a shift toward reliability-first engineering.

The introduction of multimodal memory perception and enhanced workflow variable handling opens new doors for building sophisticated, context-aware agent applications. These capabilities lay the groundwork for deeper integration between memory systems and application logic, moving MemoryBear closer to its vision of truly intelligent, adaptive memory infrastructure.

In upcoming releases, expect continued investment in workflow expressiveness, memory retrieval precision, and cross-modal understanding. The platform will deepen its support for complex agent orchestration patterns while further stabilizing the foundation for large-scale production deployments.

---

MemoryBear v0.3.0 企业版 发布说明 —— 破晓

发布日期： 2026年4月15日 ｜ 版本代号： 破晓（PoXiao · Daybreak）

MemoryBear v0.3.0 在前版基础上进行了全面升级，涵盖应用工作流、记忆智能和系统稳健性等多个维度。本版本引入了版本化API调用、多模态记忆感知以及大量工作流增强功能，同时修复了平台各层面的稳定性问题，打造更加可靠、精准且开发者友好的 MemoryBear。

---

🚀 一、核心升级概览

1. 应用与API增强

• 版本化API调用支持：对外服务API现已支持指定版本调用，使消费方可以锁定稳定接口，便于API平滑演进而不影响现有集成。
• 工作流检查清单：新增工作流检查清单功能，提供结构化的验证步骤，帮助用户在部署前核验工作流配置，减少运行时错误。
• 深度思考参数精准控制：OpenAI深度思考参数现仅发送给明确支持深度推理的模型，避免向不兼容模型注入无效参数，防止潜在错误。
• 提示器模型返回优化：优化了提示器模型的响应处理，在提示工程工作流中提供更一致、可靠的输出。

---

2. 记忆智能 🧠

• 多模态记忆感知Agent：记忆感知Agent现已支持多模态记忆的读取与写入，能够跨文本、图像等多种模态捕获更丰富的上下文，实现更全面的用户理解。
• OpenClaw内置工具：为OpenClaw新增内置工具，扩展了记忆相关Agent操作的工具集，简化集成工作流。

---

3. 用户体验 🎨

• 流式渲染稳定性优化：解决了LLM流式输出过程中页面频繁重绘导致的抖动问题，渲染管线更加平滑，模型回复时提供稳定的阅读体验。
• 记忆中枢更名：工作空间中原"记忆相关"板块已更名为"记忆中枢"，更准确地体现其作为核心记忆管理界面的定位。

---

4. 工作流改进 ⚙️

• 三级变量模板转换：工作流模板转换现已支持三级变量，修复了单文件类型节点选择无法正确解析的问题。
• VL模型Token统计：使用模型组合中的VL（视觉语言）模型时，Token用量现已正确统计，解决了与标准模型列表选择同一模型时统计缺失的差异。
• 导入工作流功能特性同步：从外部导入的工作流现已正确同步所有功能特性，包括开场白、引用和归属设置等——此前这些属性在导入时被静默丢弃。
• 会话变量名称唯一性校验：为工作流中的会话变量和开始节点变量名称新增唯一性校验，防止静默冲突和难以调试的运行时问题。
• 文件类型提取修复：工作流现可正确提取 file.type 信息，解决了此前返回空值的解析问题。
• 条件分支显示修复：条件分支节点在配置值为 0 或引用会话变量时现已正确渲染，修复了显示截断问题。
• Object/Array校验规则：为 object 和 array[object] 类型的会话变量新增校验规则，防止JSON序列化错误导致工作流无法保存。
• HTTP请求Body字段修正：工作流HTTP请求节点中，body请求体字段从使用 name 改为使用 key 进行字段标识，与标准HTTP规范对齐并解决映射问题。

---

5. 知识库 📚

• Embedding Token截断安全边界：修复了因embedding模型token超限导致的报错问题，统一为所有embedding模型添加8,000 token截断安全边界。同时优化Excel处理，每行数据保持独立chunk，不再合并，使用换行分隔以实现更清晰的文档处理。

---

6. 稳健性与缺陷修复 🔧

• 原子性更新与批量访问失败：修复了原子性记忆更新和批量访问记录操作在并发负载下可能导致数据不一致的问题。
• 对话别名提取错误：修复了应用对话中别名提取不正确的问题，确保从对话上下文中正确解析用户身份。
• 工作流别名提取：修正了工作流中的别名提取逻辑，正确区分用户和AI回复，同步考量情绪和兴趣分布（问题 #1006016）。
• RAG记忆分页：修复了RAG记忆检索中的分页数据问题，解决了跨页返回不完整或重复结果的情况。
• 隐式记忆详情显示：修复了隐式记忆条目在有对应数字值时详情中显示为空的问题（见工作流 v0.2.10 多模态记忆读取场景）。
• 向量查询驱动关闭异常：修复了记忆存储和记忆验证流程中 search_graph_by_embedding: statements 向量查询因数据库驱动提前关闭导致的异常。
• 用户管理启停异常：修复了用户管理中切换启用/停用状态时的异常API请求行为。
• 模型列表筛选不一致：修复了模型列表筛选条件与显示结果不匹配的问题，确保筛选行为准确。

---

🧭 未来展望

MemoryBear v0.3.0 标志着产品向生产成熟度迈出了坚实一步。本版本覆盖API版本化、多模态记忆、工作流稳健性和渲染稳定性等多个维度，体现了一个在持续拓展能力边界的同时快速夯实核心的平台。对边界场景的修复和开发者体验一致性的提升，标志着工程重心正向可靠性优先转变。

多模态记忆感知和增强的工作流变量处理能力的引入，为构建复杂的上下文感知Agent应用打开了新的大门。这些能力为记忆系统与应用逻辑的深度融合奠定了基础，推动 MemoryBear 向真正智能、自适应的记忆基础设施愿景迈进。

在后续版本中，我们将持续投入工作流表达力、记忆检索精度和跨模态理解能力的提升，深化对复杂Agent编排模式的支持，同时进一步稳固大规模生产部署的基础。

MemoryBear v0.2.10 Community Release Notes — Forging the Blade

Release Date: April 8, 2026 | Codename: LianJian (炼剑 · Forging the Blade)

MemoryBear v0.2.10 Community Edition delivers a comprehensive set of workflow enhancements, deeper memory capabilities, and critical stability fixes. This release introduces deep thinking mode for Agents, form-based workflow interactions, and multimodal memory reading — sharpening the platform for real-world production use.

---

🚀 I. Core Upgrade Overview

1. Workflow Engine Enhancements

• Session Variable File Support: Workflow session variables now support file-type values with configurable defaults, supporting both local and remote file sources. This enables richer data passing between workflow nodes.
• List Operation Node: A new dedicated node for list manipulation operations, allowing workflows to process, filter, and transform array data natively without custom code.
• Template Conversion HTML Support: The template conversion node now supports HTML output, expanding rendering capabilities for rich-content workflows.
• Form Return & Submission: Workflows can now return interactive form definitions to the conversation UI, and the frontend supports form submission back into the workflow — enabling structured data collection mid-conversation.
• HTTP Node XML Response: The workflow HTTP request node now supports XML format responses, broadening integration compatibility with legacy and enterprise APIs.
• Opening Remarks & File References: Workflow conversations now support configurable opening remarks and the ability to reference attached files, improving the guided conversation experience.
• Template Conversion Three-Level Variables: Template conversion nodes now support three-level nested variable access, enabling more complex data extraction from deeply structured objects.
• Node Connection Add Button: Workflow canvas node connections now include an inline add button, making it easier to insert new nodes between existing connections.

---

2. Agent Intelligence 🧠

• Agent Deep Thinking Mode: Agents now support a deep thinking mode that enables more thorough reasoning before responding. This produces higher-quality answers for complex queries at the cost of additional processing time.
• Model Deep Thinking Feature Toggle: A new model-level feature flag indicates whether a model supports deep thinking, along with a toggle to enable or disable it per application. This gives fine-grained control over reasoning behavior.

---

3. Memory System Upgrades 📚

• User Memory Pagination: The user memory library now supports pagination, improving performance and usability when browsing large memory collections.
• RAG User Memory Data Structure Refresh: The RAG user memory backend API data structures have been redesigned to align with the latest UI requirements, providing cleaner and more consistent data contracts.
• Multimodal Memory Reading: The memory system now supports reading multimodal memory entries, enabling retrieval of image, audio, and other non-text memory content.
• Semantic Pruning Threshold Hints: The memory extraction smart semantic pruning threshold now displays descriptive range labels, making it easier to understand and configure the sensitivity of pruning behavior.

---

4. Frontend & Usability 🎨

• Skill Tool Deletion Status Display: The frontend skill library tool list now displays deletion status indicators, providing clear visibility into which tools have been removed.
• Dashboard Day-over-Day Comparison: Total memory capacity, application count, knowledge base count, and API call metrics now include "compared to yesterday" data, enabling quick trend analysis on the home dashboard.

---

5. Robustness & Bug Fixes 🔧

• Parameter Extraction Null Handling: Fixed an issue where the parameter extraction node would throw exceptions when processing empty or null values. The node now handles missing data gracefully.
• Token Consumption Display Optimization: Improved the token consumption extraction and display for certain models, ensuring accurate usage reporting across different LLM providers.
• Model Parameter Negative Value Fix: Fixed unclear parameter range definitions in single-Agent model settings where negative values were ambiguously allowed or disallowed.
• App Share Deletion Sync: When a user deletes a previously shared application, the is_active field in the app_shares table now correctly updates for all share records, not just the source application entry.
• Memory Write Task Ordering: Fixed write_message task execution for the same user to respect chronological order. Single end_user tasks now execute sequentially by timestamp while multiple end_user tasks process concurrently.
• Multimodal Model Missing Graceful Handling: When no matching multimodal model is configured in memory settings, perceptual memory writes from workflow experience sharing no longer break mid-process. The system now handles the missing model gracefully.
• Custom Tool Number Variable Pass-through: Custom tools with number type parameters now correctly accept and pass through preceding variable values, resolving type coercion issues.
• Cluster Sub-Agent Display After Save: Fixed a bug where sub-agents in cluster configurations were not reflected in the UI after saving.
• Memory-Enabled Streaming Output Fix: Resolved a streaming output failure that occurred when memory was enabled, caused by a string serialization issue in the output pipeline.

---

🧭 Looking Ahead

MemoryBear v0.2.10 represents a significant step toward production maturity. The combination of deep thinking capabilities, interactive form workflows, and multimodal memory reading demonstrates the platform's evolution from a memory storage system into a comprehensive cognitive infrastructure. Each fix in this release was driven by real-world production usage, hardening the system against edge cases that only surface at scale.

The introduction of Agent deep thinking mode and the expanded workflow engine capabilities signal a clear trajectory: MemoryBear is becoming not just a memory layer, but an intelligent orchestration platform where memory, reasoning, and interaction converge. The SaaS management backend additions further position the platform for multi-tenant enterprise deployments.

We are excited for the upcoming v0.3.0 release on April 17, which will mark a major milestone for the platform. Expect deeper agent reasoning capabilities, expanded multi-agent collaboration features, and further refinements to the memory intelligence pipeline. The blade has been forged — now it's time to wield it.

---

MemoryBear v0.2.10 社区版 发布说明 —— 百炼成锋

发布日期： 2026年4月8日 ｜ 版本代号： 炼剑（LianJian · Forging the Blade）

MemoryBear v0.2.10 社区版带来了全面的工作流增强、更深层的记忆能力以及关键的稳定性修复。本版本引入了 Agent 深度思考模式、表单交互式工作流与多模态记忆读取——磨砺平台锋芒，为真实生产环境做好准备。

---

🚀 一、核心升级概览

1. 工作流引擎增强

• 会话变量文件格式支持：工作流会话变量现已支持文件类型值，可配置默认值，支持本地和远程文件源，实现节点间更丰富的数据传递。
• 列表操作节点：新增专用列表操作节点，支持工作流中原生处理、过滤和转换数组数据，无需自定义代码。
• 模板转换支持 HTML：模板转换节点现已支持 HTML 输出，扩展了富内容工作流的渲染能力。
• 表单返回与提交：工作流现可向对话界面返回交互式表单定义，前端支持将表单数据提交回工作流——实现对话中的结构化数据采集。
• HTTP 节点 XML 响应：工作流 HTTP 请求节点现已支持 XML 格式响应，拓宽了与传统及企业级 API 的集成兼容性。
• 开场白与文件引用：工作流对话现支持可配置的开场白及附件文件引用功能，提升引导式对话体验。
• 模板转换三级变量：模板转换节点现支持三级嵌套变量访问，可从深层结构化对象中提取复杂数据。
• 节点连线添加按钮：工作流画布节点连线处新增内联添加按钮，便于在已有连接之间插入新节点。

---

2. Agent 智能 🧠

• Agent 深度思考模式：Agent 现支持深度思考模式，在回复前进行更充分的推理。对于复杂查询可产出更高质量的回答，代价是额外的处理时间。
• 模型深度思考特性开关：新增模型级特性标识，标明模型是否支持深度思考，并提供应用级开关控制。实现对推理行为的精细化管理。

---

3. 记忆系统升级 📚

• 用户记忆库分页：用户记忆库现已支持分页浏览，提升大规模记忆集合的性能和可用性。
• RAG 用户记忆数据结构刷新：RAG 用户记忆后端 API 数据结构已重新设计，与最新 UI 需求对齐，提供更清晰一致的数据契约。
• 多模态记忆读取：记忆系统现支持读取多模态记忆条目，可检索图像、音频等非文本记忆内容。
• 语义剪枝阈值提示文案：记忆萃取智能语义剪枝阈值现显示描述性区间标签，便于理解和配置剪枝灵敏度。

---

4. 前端与体验 🎨

• 技能工具删除状态展示：前端技能库工具列表现显示删除状态标识，清晰展示哪些工具已被移除。
• 仪表盘日环比数据：总记忆容量、应用数量、知识库数量及 API 调用次数指标现包含"与昨日相比"数据，支持首页仪表盘的快速趋势分析。

---

5. 稳健性与缺陷修复 🔧

• 参数提取空值处理：修复参数提取节点在处理空值或 null 值时抛出异常的问题，现已优雅处理缺失数据。
• Token 消耗展示优化：优化部分模型的 token 消耗提取与展示，确保不同 LLM 提供商的用量报告准确。
• 模型参数负值修复：修复单 Agent 模型参数设置中参数范围定义不明确、负值允许与否模糊的问题。
• 应用共享删除同步：用户删除已共享应用后，app_shares 表中的 is_active 字段现可正确更新所有共享记录，而非仅更新源应用条目。
• 记忆写入任务排序：修复同一用户的 write_message 任务执行顺序问题。单 end_user 任务现按时间戳顺序执行，多 end_user 任务并发处理。
• 多模态模型缺失优雅处理：当记忆配置中未配置对应的多模态模型时，工作流体验分享的感知记忆写入不再中断，系统现已优雅处理模型缺失情况。
• 自定义工具 Number 变量传递：number 类型参数的自定义工具现可正确接受和传递前置变量值，解决类型转换问题。
• 集群子代理保存后显示：修复集群配置中子代理保存后未在界面反显的问题。
• 记忆开启后流式输出修复：解决开启记忆后流式输出失败的问题，原因为输出管道中的字符串序列化问题。

---

🧭 未来展望

MemoryBear v0.2.10 标志着平台向生产成熟度迈出的重要一步。深度思考能力、交互式表单工作流与多模态记忆读取的结合，展现了平台从记忆存储系统向综合认知基础设施的演进。本版本的每一项修复都源于真实生产环境的使用反馈，使系统在规模化运行中更加坚韧。

Agent 深度思考模式的引入与工作流引擎能力的扩展，指明了清晰的发展方向：MemoryBear 正在成为不仅仅是记忆层，而是记忆、推理与交互融合的智能编排平台。SaaS 管理后台的新增功能进一步为多租户企业级部署奠定基础。

我们期待 4 月 17 日即将到来的 v0.3.0 发布会，这将是平台的一个重要里程碑。届时将带来更深层的 Agent 推理能力、扩展的多智能体协作功能，以及记忆智能管线的进一步优化。剑已炼成，只待出鞘。

MemoryBear v0.2.9 Community Release Notes — Listening to the Wind

MemoryBear v0.2.9 Community Release Notes — Listening to the Wind

Release Date: March 31, 2026 | Codename: TingFeng (听风 · Listening to the Wind)

MemoryBear v0.2.9 Community delivers a deep upgrade across application capabilities and memory intelligence. This release introduces application logging, advanced episodic memory retrieval with community-based search, multimodal memory persistence into Neo4j, and user alias management — bringing the platform closer to its v0.3.0 launch milestone with refined observability and richer cognitive recall.

---

🚀 I. Core Upgrade Overview

1. Application Framework

• Model-Feature Binding: Model selection is now bound to application features such as file upload types, ensuring that only compatible models are available for each capability.
• Application Logging (Message Records): A new application log system captures full message records across Agent, Workflow, and Agent Cluster interactions. Sharers can view experience sharing logs and API Key call logs (using their own key); shared users can view API Key call logs (using their own key).
• Application Features — Conversation Opener & Document Citation: Applications now support configurable conversation openers (opening remarks) and document citation in replies, enriching the user interaction experience.
• API Key Search: Added the ability to search applications by API Key, enabling quick lookup and management of key-to-app associations.
• Agent Memory Defaults: Agents now have "conversation history memory" enabled by default. When shared via experience sharing, "memory function" is enabled by default and can be toggled off — but only if "conversation history memory" was already enabled; otherwise, the memory function is not available on the shared page.
• Application Copy Fix: Fixed an issue where modifying a copied application incorrectly sent the original app's app_id instead of the copied app's app_id.

---

2. Workflow Enhancements 🔧

• Document Extraction Node: A new workflow node for extracting structured content from documents, enabling document-driven automation pipelines.
• Reply Node Output Continuation: Reply node outputs can now connect to additional downstream nodes, allowing post-reply processing and branching.
• Workflow Memory Write with File Variables: Workflow memory write nodes now support configuring file-type variables via sys.files, enabling multimodal content to be persisted into Neo4j.
• Conditional Branch Whitespace Fix: Fixed an issue where conditional branch nodes failed when conditions contained spaces or newline characters.

---

3. Memory Intelligence 🧠

• Multimodal Memory into Neo4j: Multimodal memory (voice, image, video) can now be persisted into Neo4j. Memory configuration supports additional model selection for voice, image, and video with on/off toggles. In this release, only Workflow-based multimodal memory write is supported; Agent support is planned for the next version. Perceptual memory must be associated with a multimodal model and a memory node to be written. Neo4j storage adds a new perceptual memory node type linked to Chunk nodes. Memory engine configuration now includes model filtering by capability — visual, audio, and video — requiring selection of models with matching abilities.
• Episodic Memory Retrieval with Community Search: Episodic memory retrieval now includes community-based search results across three retrieval modes: Quick Reply returns summaries (community summaries); Normal Reply returns summaries + communities (topic structures) + statements (atomic facts/evidence); Deep Thinking returns summaries + communities + statements + expanded statements (graph-extended implicit evidence).
• User Alias Management: Agent conversations now extract user name/alias information and display it in the "Core Profile." Editing and saving aliases in the Core Profile syncs to the end_user_info table's other_name field. Multiple aliases are supported with time-based extraction priority. Alias data is accessible via API.
• Community Node Summary Enrichment: Community node summaries are now enriched with entity_name, entity_description, and Statement.statement content, providing more comprehensive community-level context.
• Pilot Run Progress Callback: In pilot run (draft run) mode, a progress callback event is now sent upon knowledge extraction completion, including total statement count and chunk count in the payload.
• Knowledge Extraction Complete Event: A dedicated knowledge_extraction_complete event is now emitted after the knowledge_extraction_result phase concludes, providing a clear lifecycle signal for downstream consumers.
• Memory Cache System: Added caching for memory system operations, improving retrieval performance and reducing redundant computation.
• Semantic Pruning Optimization: Fixed an issue where semantic pruning during memory extraction incorrectly removed emotion and interest/hobby data. Improved pruning behavior across multiple ontology scenarios to produce differentiated results.
• Forgetting Cycle Task Fix: Fixed the run_forgetting_cycle_task scheduled task that was failing during forgetting cycle execution.
• Memory Config ID Type Fix: Fixed an issue where applications using integer-type config_id values failed to use the selected memory configuration, falling back to the workspace default config instead.
• Memory Write Issues: Fixed structured output failures and corrected clustering sequence to occur before memory write operations.
• User Memory Config Version Fix: Fixed an issue where user memory configuration was not correctly mapped to the published version's config record.

---

4. Model Management 🤖

• Volcengine Model Provider: Added Volcengine (火山引擎) as a new model provider, expanding the range of available LLM backends.
• Custom Model Capabilities: Custom models now support declaring full-modality, visual, audio, and video capabilities, enabling precise model-feature matching.
• API Key Masking: Model list API responses now mask API keys, showing only the first and last 4 characters with **** in between, improving security posture.

---

5. User & Platform 👥

• Username Uniqueness Relaxed: Usernames no longer require uniqueness — only email addresses must be unique, simplifying user registration and management.
• Frontend UI Validation: Comprehensive frontend UI validation pass across application pages (page checklist pending from frontend team).

---

6. Robustness & Bug Fixes 🔧

• Remote File Document Type Support: Multimodal remote files now support document types in addition to media files.
• File Metadata Location: Fixed inconsistency in meta_data where file references were stored under both files and files[] formats.
• Text-to-Speech Status Feedback: Backend now provides generation status for text-to-speech, keeping the speaker icon in "generating" state until audio is ready for playback.
• Model Capability Feature Binding: Fixed model capabilities in Agent, Workflow, and Cluster not properly linking to application features.
• Workflow Loop Node Termination Condition: Fixed loop node "missing right field" error when setting non-empty termination conditions.
• MCP Market Service Count Reset: Fixed sidebar "ModelScope MCP" showing 9295 available services on first click, then 0 on subsequent clicks.
• Dashboard App Count with Shared Apps: The dashboard_data endpoint now correctly includes shared applications in the total app count.
• Agent Web Search Recursion Limit: Fixed Agent web search triggering "Recursion limit of 9 reached without hitting a stop condition."
• RAG Workspace Workflow Memory Write: Fixed workflows in RAG workspaces unable to write to the memory store.
• Tool Market Logo Mismatch: Fixed incorrect tool vendor logos in the tool market configuration view.
• TTS Streaming Failure on Long Text: Fixed text-to-speech streaming failure when conversation text exceeded length thresholds.
• Shared App Conversation Summary: Fixed shared app conversations with openers always displaying "New conversation" as the sidebar summary.
• Perceptual Memory Text Preview: Fixed preview and download issues for text-based perceptual memory entries.
• MCP Market Unauthorized Error: Fixed mcp_market_configs/operational_mcp_servers endpoint returning 401 Unauthorized.
• Legacy RAG Workspace Agent Issues: Fixed Agent conversation anomalies in legacy RAG workspaces.
• Shared App Experience Issues: Fixed historical conversation TTS playback and conversation parameter passing in shared app experiences.

---

🧭 Looking Ahead

By deepening multimodal memory persistence, introducing community-based episodic retrieval, and expanding application observability through logging, this release solidifies the cognitive foundation that MemoryBear is built upon — perception, refinement, association, and forgetting working in concert.

The introduction of community search in episodic memory retrieval marks a qualitative leap in how MemoryBear surfaces knowledge. With three distinct retrieval depths — from quick summaries to graph-expanded implicit evidence — the platform now offers a retrieval experience that mirrors human cognitive recall: fast intuition for simple queries, structured reasoning for complex ones.

v0.3.0 will mark MemoryBear's official product launch event, bringing together all the capabilities refined across the 0.2.x series into a cohesive, production-ready platform. We look forward to unveiling the full vision of MemoryBear as the cognitive memory layer for AI applications at the launch conference.

---

MemoryBear v0.2.9 社区版 发布说明 —— 听风知意

发布日期： 2026年3月31日 ｜ 版本代号： 听风（TingFeng · Listening to t...

MemoryBear v0.2.8 Community Release Notes — Jade in Splendor

MemoryBear v0.2.8 Community Release Notes — Jade in Splendor

Release Date: March 20, 2026 | Codename: JingYu (景玉 · Jade in Splendor)

MemoryBear v0.2.8 Community brings major upgrades to application sharing, multimodal interaction, and platform infrastructure. This release introduces voice input/output, multimodal perceptual memory, cloud file storage, and workspace-level identity — delivering a more capable and resilient open platform for AI memory management.

---

🚀 I. Core Upgrade Overview

1. Application Sharing & Publishing

• Application Sharing (Agent, Workflow, Agent Cluster): Full sharing support across all application types — Agents, Workflows, and Agent Clusters can now be shared to other workspaces.
• Memory Enabled by Default on Shared Apps: When an application is published and shared, memory is enabled by default. Users receive a reminder when attempting to disable it, ensuring conversational context continuity.
• Workflow Memory Sharing Rules: Workflows configured with memory extraction and memory storage have memory enabled by default on the experience sharing page and cannot be disabled; workflows without memory configuration have memory disabled by default and cannot be enabled.
• Shared Session Web Search Fix: Fixed an issue where web search functionality stopped working after sharing an application session, restoring full internet access for shared apps.

---

2. Multimodal & Interaction 💬

• Voice Input for Models & Apps: Model interfaces and applications now support voice input, enabling hands-free interaction with Agents and Workflows.
• Rich Reply Capabilities: Applications now support voice reply modality — significantly expanding how AI can deliver results to users.
• Multimodal Memory — Perceptual Memory: The memory system now supports perceptual memory for multimodal inputs, enabling the platform to perceive and remember visual, audio, image, and file interactions across Agent dry runs, experience sharing, and API calls.
• File Display in Chat (Debug + Share): Uploaded files, images, videos, and audio now display correctly in both Agent and Workflow dry-run and experience sharing chat interfaces.

---

3. Platform & Infrastructure ⚙️

• i18n Internationalization: Full internationalization support, enabling the platform to serve multi-language, multi-region users.
• Cloud File Storage (OSS + S3): The file system now supports cloud uploads via Alibaba Cloud OSS or S3, ensuring reliability and scalability for all documents.
• Flower Container for Celery Monitoring: A new Flower container for monitoring and managing Celery async and scheduled task execution. Each environment runs 3 workers, and API service restarts have been verified not to cause task interruption or log loss.

---

4. EndUser Identity Migration 🔐

• EndUser Migration from app_id to workspace_id: EndUser identity has been migrated from application-level to workspace-level. Dry runs use the logged-in user's user_id as the EndUser other_id (unique per workspace); experience sharing passes other_id from the frontend; API+Key calls use the third-party ID.

---

5. Episodic Memory 🧠

• Episodic Memory Clustering Algorithm: Implemented a community-graph-based episodic memory clustering algorithm, including support for generating community graphs for existing users.

---

6. Robustness & Bug Fixes 🔧

• MCP Service Deletion Tool 404 Error: Fixed an issue where removing an MCP service from an Agent caused the api/tools endpoint to return a "404, tool does not exist" error.
• App Export Canvas vs. Saved Config Mismatch: Copying or exporting an Agent/Workflow via the top-left menu now correctly exports the saved configuration rather than the current unsaved canvas state.
• Workflow Duplicate Node ID Error: Fixed an issue where duplicating a workflow node triggered "Invalid workflow config: node IDs must be unique, duplicate ID: llm_qa".
• Conditional Branch Wiring Error: Fixed an issue where after adding a conditional branch/question classifier node, IF (case0) and ELSE (case1) connected to different downstream nodes, but after saving and refreshing both downstream nodes connected to IF (case0).
• Reply Node Content Loss: Fixed an issue where clicking reply node details and then clicking the canvas caused reply node content to disappear.
• Workflow Port Connection Rules: Optimized workflow node connection ports — left-to-left and right-to-right connections are now prohibited, and the end node option is no longer shown when adding nodes from left-side ports.
• Knowledge Base List Status Column Width: The status column in the knowledge base list now has a locked or adaptive width, preventing layout issues.
• Pending Document Preview Support: Documents in "pending" status (parsing enabled but not yet complete) now support preview for all file types.
• Knowledge Base Association Fixes: Consolidated fixes for multiple issues when associating knowledge bases with applications.
• Multimodal Conversation Continuity: Fixed an issue where sending multimodal content (images, videos, files, audio) in an Agent or Workflow only allowed one round of conversation, preventing follow-up questions based on uploaded media.
• Memory Storage Timezone Unification: Database storage time and task execution time now use a unified timezone controlled by environment variables, eliminating cross-environment inconsistencies.
• Forgetting Strength Decimal Precision: Fixed an issue where memory forgetting strength values displayed excessively long decimals.

---

🧭 Looking Ahead

MemoryBear v0.2.8 represents a significant step toward enterprise production readiness. The introduction of SaaS administration, independent service deployment, and workspace-level identity architecture demonstrates the platform's evolution from a development tool into a governed, scalable infrastructure layer for AI memory management.

The multimodal memory capabilities introduced in this release mark a pivotal expansion of what MemoryBear can perceive and retain. By enabling perceptual memory across visual, audio, and file interactions, the platform moves closer to truly human-like cognitive assistance — where context is not limited to text but encompasses the full spectrum of user interaction.

In upcoming releases, we will deepen multi-agent collaboration patterns, expand the episodic memory clustering framework, and continue hardening the platform for high-availability enterprise deployments. Enhanced analytics, fine-grained permission controls, and broader MCP ecosystem integration are also on the horizon.

---

MemoryBear v0.2.8 社区版 发布说明 —— 景玉生辉

发布日期： 2026年3月20日 ｜ 版本代号： 景玉（JingYu · Jade in Splendor）

MemoryBear v0.2.8 社区版在应用共享、多模态交互和平台基础设施方面带来重大升级。本版本引入语音输入输出、多模态感知记忆、云端文件存储和工作空间级身份体系，为 AI 记忆管理提供更强大、更稳健的开放平台。

---

🚀 一、核心升级概览

1. 应用共享与发布

• 应用共享（Agent、工作流、Agent 集群）：全面支持所有应用类型的共享——Agent、工作流和 Agent 集群均可共享至其他空间。
• 分享应用默认开启记忆功能：应用发布分享后，记忆功能默认开启。用户尝试关闭时会收到提醒，确保上下文对话的连续性。
• 工作流记忆分享规则：工作流配置了记忆提取和记忆存储后，体验分享页面默认开启记忆功能且不可取消；未配置记忆的工作流则默认关闭记忆功能且不可开启。
• 分享会话联网搜索修复：修复了应用会话分享后联网搜索功能失效的问题，恢复分享应用的完整联网能力。

---

2. 多模态与交互 💬

• 模型与应用支持语音输入：模型接口和应用现已支持语音输入，实现 Agent 和工作流的免手操作交互。
• 丰富的回复能力：应用现支持语音回复模态——大幅拓展了 AI 传达结果的方式。
• 多模态记忆——记忆库感知记忆：记忆系统现支持多模态输入的感知记忆，使平台能够在 Agent 试运行、体验分享和 API 调用中感知和记忆视觉、音频、图片和文件交互。
• 对话框中展示上传文件（调试 + 分享）：上传的文件、图片、视频和音频现可在 Agent 和工作流的试运行及体验分享对话界面中正确展示。

---

3. 平台与基础设施 ⚙️

• i18n 国际化：全面支持国际化，使平台能够服务多语言、多地区用户。
• 云端文件存储（OSS + S3）：文件系统现支持通过阿里云 OSS 或 S3 上传至云端，确保所有文档的可靠性和可扩展性。
• Flower 容器监控 Celery 任务：新增 Flower 容器用于监控和管理 Celery 异步任务及定时任务的执行状态。每个环境运行 3 个 worker，已验证 API 服务重启不会导致任务中断或日志丢失。

---

4. EndUser 身份迁移 🔐

• EndUser 从 app_id 迁移至 workspace_id：End_User 身份从应用级迁移至工作空间级。试运行取登录用户的 user_id 作为 End_User 的 other_id（同一空间唯一）；体验分享由前端传递 other_id；API+Key 调用取第三方 ID。

---

5. 情景记忆 🧠

• 情景记忆聚类算法：实现了基于社区图谱的情景记忆聚类算法，包括为已有老用户生成社区图谱的支持。

---

6. 稳健性与缺陷修复 🔧

• 删除 MCP 服务后工具 404 错误：修复了从 Agent 中删除 MCP 服务后，api/tools 接口返回"404，工具不存在"错误的问题。
• 应用导出保存画布与已保存配置不一致：通过左上角复制或导出 Agent/工作流时，现在正确导出已保存的配置而非当前未保存的画布状态。
• 工作流复制节点 ID 重复错误：修复了复制工作流节点后运行提示"工作流配置无效：节点 ID 必须唯一，重复的 ID: llm_qa"的问题。
• 条件分支连线错误：修复了添加条件分支/问题分类器节点后，IF（case0）和 ELSE（case1）分别连接不同下游节点，保存刷新后两个下游节点都连到 IF（case0）的问题。
• 回复节点内容丢失：修复了点击回复节点详情后再点击画布，回复节点内容消失的问题。
• 工作流连接桩规则优化：优化了工作流节点连接桩——禁止左左相连和右右相连，左侧连接桩添加节点时不再展示结束节点选项。
• 知识库列表状态列宽度：知识库列表的状态列现已锁定宽度或自适应，防止布局错乱。
• 等待中文档支持预览：处于"等待中"状态（已开启解析但未完成）的文档现支持各类型文件的预览。
• 知识库关联问题汇总修复：统一修复了应用中关联知识库时的多个问题。
• 多模态对话连续性：修复了在 Agent 或工作流中发送多模态内容（图片、视频、文件、音频）后只能进行一轮对话、无法基于已上传媒体继续提问的问题。
• 记忆存储时区统一：数据库存储时间和任务执行时间现采用环境变量统一控制时区，消除跨环境不一致问题。
• 遗忘强度小数精度：修复了记忆遗忘强度值显示过长小数的问题。

---

🧭 未来展望

MemoryBear v0.2.8 标志着平台向企业生产就绪迈出的重要一步。SaaS 管理后台、独立服务部署和工作空间级身份架构的引入，展现了平台从开发工具向可治理、可扩展的 AI 记忆基础设施层的演进。

本版本引入的多模态记忆能力标志着 MemoryBear 感知与留存能力的关键性扩展。通过在视觉、音频和文件交互中启用感知记忆，平台正朝着真正类人的认知辅助方向迈进——上下文不再局限于文本，而是涵盖用户交互的全部维度。

在后续版本中，我们将深化多智能体协作模式、扩展情景记忆聚类框架，并持续强化平台的高可用企业部署能力。增强的分析能力、细粒度权限控制和更广泛的 MCP 生态集成也已提上日程。

MemoryBear v0.2.7 Release Notes

MemoryBear v0.2.7 Release Notes

Release Date: March 13, 2026 | Codename: WuLing (武陵 · Wuling)

MemoryBear v0.2.7-community marks a significant milestone in platform maturity, bringing powerful capabilities for application portability, tool ecosystem expansion, and memory intelligence refinement. This release introduces application import/export for agents and workflows, integrates the MCP Marketplace for tool discovery, and delivers critical improvements to implicit and emotional memory generation logic. Alongside workflow enhancements, v0.2.7-community strengthens the foundation for production-ready deployments with comprehensive bug fixes addressing RAG space memory generation, semantic pruning accuracy, and end-user management.

---

🚀 I. Core Upgrade Overview

1. Application Management & Portability

• Application Import/Export: Full support for importing and exporting applications including both agent configurations and workflow definitions, enabling seamless migration, backup, and sharing of complex application setups across environments.

---

2. Tool Ecosystem Expansion 🔌

• MCP Marketplace Integration: Tool management now includes access to the MCP Marketplace, providing a centralized hub for discovering, browsing, and integrating Model Context Protocol tools directly within the platform.

---

3. Workflow Enhancements 📝

• Annotation Node: Introduced a new annotation node type in workflow builder, allowing developers to add inline documentation, comments, and contextual notes directly within workflow graphs for improved maintainability and team collaboration.

---

4. Memory Intelligence Refinement 🧠

• Implicit & Emotional Memory Generation Logic: Comprehensive optimization of implicit memory and emotional memory generation pipelines, including data existence validation to prevent duplicate processing, timeline-based filtering for scheduled tasks to ensure temporal accuracy, and cache validation for interest distribution to improve consistency and reduce redundant computations.
• Interest Distribution Generation Logic: Refined the interest distribution generation algorithm to produce more accurate and nuanced user interest profiles, enhancing personalization capabilities across memory-driven features.

---

5. User Experience Improvements 🎨

• Knowledge Base Sharing Loading State: Added loading indicators to knowledge base sharing operations, providing clear visual feedback during share link generation and improving perceived responsiveness.

---

6. Robustness & Bug Fixes 🔧

• End User Management in App Debugging: Fixed issue where application debugging sessions incorrectly created or associated end_user records, ensuring proper user context isolation during development workflows.
• Knowledge Base Dataset Creation Flow: Resolved bug preventing users from proceeding to the next step after creating a dataset in knowledge base setup, restoring smooth onboarding experience.
• RAG Space Memory Generation Failure: Fixed critical issue where memory generation failed and storage was interrupted in RAG-enabled spaces, ensuring reliable memory persistence across all workspace configurations.
• Application Character Limit Enforcement: Added conditional validation to application name and description fields, preventing edge cases with excessively long input that could cause UI rendering or database storage issues.
• Semantic Pruning Emotion/Interest Preservation: Optimized memory extraction semantic pruning logic to prevent incorrect deletion of emotional and interest-related memory fragments, preserving critical affective and preference data during consolidation.
• Semantic Pruning Effectiveness: Enhanced the overall effectiveness of semantic pruning algorithms, improving the balance between memory compression and information retention for more intelligent long-term memory management.

---

🧭 Looking Ahead

Version 0.2.7-community represents a maturation of MemoryBear's core infrastructure, moving from foundational capabilities to production-ready features. The introduction of application portability and MCP Marketplace integration signals a shift toward ecosystem thinking—enabling not just powerful individual deployments, but interconnected, shareable, and extensible memory-driven systems. The refinements to implicit and emotional memory generation demonstrate a commitment to measurable intelligence quality as the platform scales.

As we continue to refine memory intelligence algorithms, the next phase will focus on memory retrieval benchmarking and episodic memory clustering algorithms to enhance contextual recall and temporal reasoning capabilities.

Looking forward to v0.2.8 and beyond, we will introduce multimodal memory capabilities with distributed service deployment for knowledge bases and models, enabling voice input for applications and expanding application capabilities with voice responses, BI visualizations, PPT generation, and direct image creation. Application conversation sharing and web search functionality will be restored and enhanced. The journey toward truly intelligent, multimodal, context-aware applications continues.

---

MemoryBear v0.2.7 发布说明

发布日期： 2026年3月13日 ｜ 版本代号： 武陵（WuLing · Wuling）

MemoryBear v0.2.7-community 标志着平台成熟度的重要里程碑，带来应用可移植性、工具生态扩展和记忆智能精细化能力。本版本引入了 Agent 和工作流的应用导入/导出功能，集成 MCP 广场实现工具发现，并对隐性记忆和情绪记忆生成逻辑进行了关键改进。伴随工作流增强，v0.2.7-community 通过全面的缺陷修复（涵盖 RAG 空间记忆生成、语义剪枝准确性和终端用户管理）夯实了生产就绪部署的基础。

---

🚀 一、核心升级概览

1. 应用管理与可移植性

• 应用导入/导出：全面支持应用的导入和导出，包括 Agent 配置和工作流定义，实现复杂应用设置在不同环境间的无缝迁移、备份和共享。

---

2. 工具生态扩展 🔌

• MCP 广场集成：工具管理现已接入 MCP 广场，提供集中式枢纽用于发现、浏览和集成模型上下文协议（MCP）工具，直接在平台内完成工具生态对接。

---

3. 工作流增强 📝

• 备注节点：在工作流构建器中引入全新的备注节点类型，允许开发者在工作流图中直接添加内联文档、注释和上下文说明，提升可维护性和团队协作效率。

---

4. 记忆智能精细化 🧠

• 隐性记忆与情绪记忆生成逻辑优化：全面优化隐性记忆和情绪记忆生成管线，包括数据存在性校验以防止重复处理、定时任务的时间轴筛选以确保时序准确性，以及兴趣分布的缓存校验以提升一致性并减少冗余计算。
• 兴趣分布生成逻辑改进：优化兴趣分布生成算法，产生更准确、更细腻的用户兴趣画像，增强记忆驱动功能的个性化能力。

---

5. 用户体验改进 🎨

• 知识库分享加载状态：为知识库分享操作增加加载指示器，在分享链接生成过程中提供清晰的视觉反馈，改善感知响应速度。

---

6. 稳健性与缺陷修复 🔧

• 应用调试中的终端用户管理：修复应用调试会话错误创建或关联 end_user 记录的问题，确保开发工作流中的用户上下文正确隔离。
• 知识库数据集创建流程：解决知识库设置中创建数据集后无法进入下一步的缺陷，恢复流畅的引导体验。
• RAG 空间记忆生成失败：修复 RAG 启用空间中记忆生成失败和存储中断的关键问题，确保所有工作空间配置下的可靠记忆持久化。
• 应用字符限制强制执行：为应用名称和描述字段增加条件校验，防止过长输入导致 UI 渲染或数据库存储问题的边界情况。
• 语义剪枝情绪/兴趣保留：优化记忆萃取语义剪枝逻辑，防止错误删除情绪和兴趣相关的记忆片段，在整合过程中保留关键的情感和偏好数据。
• 语义剪枝效果优化：增强语义剪枝算法的整体有效性，改善记忆压缩与信息保留之间的平衡，实现更智能的长期记忆管理。

---

🧭 未来展望

v0.2.7 版本代表了 MemoryBear 核心基础设施的成熟，从基础能力迈向生产就绪特性。应用可移植性和 MCP 广场集成的引入标志着向生态思维的转变——不仅支持强大的独立部署，更实现互联、可共享和可扩展的记忆驱动系统。隐性记忆和情绪记忆生成的精细化改进展现了平台在规模化过程中对可衡量智能质量的承诺。

随着记忆智能算法的持续优化，下一阶段将聚焦记忆检索基准测试和情景记忆聚类算法，增强上下文召回和时序推理能力。

展望 v0.2.8 及更远的未来，我们将引入多模态记忆能力，实现知识库和模型的分服务部署，为应用增加语音输入支持，并扩展应用能力至语音回复、BI 可视化、PPT 生成和直接生图。应用会话分享和联网搜索功能将得到修复和增强。通往真正智能、多模态、上下文感知应用的旅程仍在继续。