Skip to content

MemoryBear v0.3.6 Community Release Notes — Converging Currents

Choose a tag to compare

View all tags
@keeees keeees released this 12 Jun 03:11
· 3990 commits to main since this release
v0.3.6
ab8ac7e

Release Date: June 3, 2026 | Codename: HuiLan (汇澜 · Converging Currents)

MemoryBear v0.3.6 Community builds on the sharpening work of v0.3.5 by widening the flow of every major surface simultaneously. Workflow orchestration gains trigger nodes, cache management, error handling, and a richer start-node model. Entity-level reflection now supports description-summary merging, rename suggestions, and dual-track entity deduplication, while memory-assisted answering gains support for hypothetical questions. The knowledge base infrastructure is hardened with parse-time resilience, chunk-size flexibility, and improved retrieval reliability. Together, these currents converge into a platform that is more autonomous, more resilient, and ready for production use at scale.

🚀 I. Core Upgrade Overview

1. Workflow & Application Orchestration ⚙️

  • Dify HTTP Node — Null Auth Header Compatibility: Workflows imported from Dify that include HTTP nodes with authentication now correctly handle the extra header parameter added during import; passing null no longer causes MemoryBear to throw an error, ensuring compatibility across all Dify export shapes.
  • Start Node — Next Step & Parallel Support: The workflow start node now supports triggering the next step and branching into parallel execution paths, unlocking more complex orchestration patterns directly from the entry point.
  • New Trigger Node: A dedicated trigger node has been added to the workflow canvas, enabling event-driven and scheduled workflow initiations without wrapping logic inside the start node.
  • Workflow Published as Tool — Enhanced Capabilities: When a workflow is published as a tool, it now exposes three additional built-in capabilities: suggested follow-up questions, bidirectional voice transcription, and content moderation — giving tool consumers richer interaction patterns out of the box.
  • Iterative Node Error Handling: The iterate node now supports configurable error handling, allowing workflows to define fallback behavior when an iteration step fails rather than halting the entire run.
  • Workflow Cache Management: A new cache management panel lets builders view and control the caching layer within workflows, making it easier to reason about stale state and force-refresh during development and debugging.
  • Code Node — Failure Handling: The code execution node gains a dedicated failure-handling path, so downstream nodes can respond gracefully to code errors instead of receiving an unhandled exception propagation.
  • Start Node — Waiting State: The workflow start node now supports a waiting state, allowing the workflow to pause at entry until a condition or external signal is satisfied before continuing.

2. Memory Intelligence 🧠

  • Answer Citing Memory — Precision Fix: Fixed a pipeline regression where memory-assisted answers were no longer influencing response precision. The memory reference path is now correctly wired into the answer pipeline, restoring the accuracy improvement that cited memory is expected to deliver.
  • Pipeline Snapshot Output Refactor: The pipeline snapshot output structure has been refactored for cleaner formatting and more predictable downstream consumption, reducing integration friction for teams that inspect or store pipeline traces.
  • Memory Graph Style Update: The memory knowledge graph UI has been updated; when the node count exceeds 1,000 the view applies an automatic filter to keep the graph navigable, preventing render degradation on dense workspaces.
  • Entity Reflection — Summary & Rename Suggestion: The reflection step now merges new entity description fragments into a cleaner description_summary and suggests clearer entity names when the evidence supports it.
  • Dual-Track Entity Deduplication: Entity deduplication now combines name semantic-similarity matching with same-type entity deduplication, reducing duplicate entity accumulation from both naming variants and type-level overlap.
  • Memory-Assisted Hypothetical Question Answering: When answering user questions, MemoryBear can now reason over hypothetical scenarios using memory context to construct plausible answers grounded in stored information.
  • Homepage Memory Tags: A new memory tag surface has been added to the homepage, giving users a quick visual index into the active memory clusters for the current workspace.

3. Knowledge Base 📚

  • MCP Service — CSV Import Content Completeness Fix: Fixed a bug in the MCP service's CSV import path where not all rows were being written to the knowledge base; imports now correctly ingest the full dataset.
  • Feishu Knowledge Base — 404 File Preview/Download: Fixed an issue where Feishu knowledge base files that return a 404 response were blocking preview and download flows; the error is now handled gracefully with appropriate user feedback.
  • Chunk Size Range Extended to 6–12,000 Characters (Parent-Child Strategy): The allowed chunk size range for knowledge base documents using the parent-child chunking strategy has been expanded from the previous ceiling to support up to 12,000 characters, enabling higher-fidelity context windows for long-form document retrieval.
  • Hybrid & Graph Retrieval — Rerank Error Fix: Fixed a runtime error (Failed to rerank documents: 'detail') that occurred during hybrid retrieval and graph retrieval, preventing reranking from completing and degrading result quality.
  • Document Deletion During Parse — Task Release Fix: Fixed a critical task management bug where deleting a document while it was still being parsed left the task running in the document queue indefinitely; the parse task now detects the deletion signal and releases its slot, allowing queued tasks to proceed.
  • Recall Test — Hybrid Mode Missing Count Error Fix: Fixed a bug where running a recall test in the default hybrid retrieval mode without explicitly setting a recall count produced an error; the count now defaults gracefully and the test completes.

4. Robustness & Task Management 🔧

  • Celery Inspection Task — Mutex Lock on Trigger: The Celery inspection task now sets a mutual exclusion lock when triggered, preventing concurrent duplicate executions from competing over the same inspection window and causing inconsistent state.

🧭 Looking Ahead

MemoryBear v0.3.6 Community marks the point where the platform begins to feel like a coherent whole rather than a collection of capable parts. The workflow layer now has the vocabulary it needs for real orchestration: triggers, parallel branches, error paths, cache visibility, and a waiting start node. The memory layer now has a clearer entity-level reflection path for summary merging, rename suggestions, and dual-track entity deduplication. The knowledge base has the resilience it needs for real data pipelines: parse-time task management and expanded chunk sizes. These are not incremental additions — they are the connective tissue that turns individual capabilities into a production-grade system.

The memory-engine change is focused on making entity state easier to maintain safely. New description fragments can be compacted into stable summaries, duplicate entities can be reduced through dual-track deduplication, and overly generic entity names can be improved when the evidence is clear.

MemoryBear v0.3.6 社区版 发布说明 —— 汇流成澜

发布日期： 2026年6月3日 ｜ 版本代号： 汇澜（HuiLan · Converging Currents）

MemoryBear v0.3.6 社区版在 v0.3.5 精打细磨的基础上，同时拓宽了每条主要脉络的流量。工作流编排新增触发器节点、缓存管理、错误处理机制，以及更丰富的开始节点模型；实体级反思支持描述摘要合并、实体重命名建议和双轨实体去重，记忆辅助回答支持假设性问题；知识库基础设施在解析韧性、分块灵活度和检索可靠性上全面加固。诸流汇聚，形成一套更自主、更健壮、面向规模化生产使用的平台。

🚀 一、核心升级概览

1. 工作流与应用编排 ⚙️

  • Dify HTTP 节点 —— Null 鉴权 Header 兼容：修复从 Dify 导入的工作流中 HTTP 节点鉴权多出一个 header 参数的问题；传入 null 时 MemoryBear 不再报错，兼容所有 Dify 导出格式。
  • 开始节点 —— 支持下一步与并行：工作流开始节点现在支持触发"下一步"以及分支至并行执行路径，直接从入口点解锁更复杂的编排模式。
  • 新增触发器节点：工作流画布中新增独立触发器节点，支持事件驱动与定时启动工作流，无需将触发逻辑嵌入开始节点。
  • 工作流发布成工具 —— 增强功能：工作流发布为工具后，额外开放三项内置能力：建议问题、语音互转、内容审查，让工具消费方开箱即获得更丰富的交互模式。
  • 迭代节点支持错误处理：迭代节点支持配置错误处理路径，工作流可在某步迭代失败时定义兜底行为，不再因单步失败而中断整体运行。
  • 工作流新增缓存管理：新增缓存管理面板，支持在工作流内查看和控制缓存层，便于在开发与调试期间排查陈旧状态并手动刷新。
  • 代码节点支持失败处理：代码执行节点新增独立的失败处理路径，下游节点可对代码错误进行优雅响应，不再接收未经处理的异常传播。
  • 开始节点加上等待状态：工作流开始节点支持等待状态，工作流可在入口处暂停，直至满足条件或外部信号触发后才继续执行。

2. 记忆智能 🧠

  • 回答引用记忆 —— 精准度修复：修复流水线回归问题，记忆辅助回答的精准度提升未能生效；记忆引用路径已重新正确接入回答流水线，恢复引用记忆本应带来的准确性提升。
  • 流水线快照输出改造：流水线快照输出结构经过重构，格式更整洁，下游消费更可预期，降低检查或存储流水线追踪记录团队的集成摩擦。
  • 记忆图谱样式更新：记忆知识图谱 UI 已更新；当节点数量超过 1000 时，视图自动筛选以保持图谱可导航，防止稠密工作区出现渲染退化。
  • 实体反思 —— 摘要与重命名建议：反思步骤现在可以将新的实体 description 碎片合并为更精炼的 description_summary，并在证据清楚时建议更稳定的实体名称。
  • 双轨实体去重：实体去重现在同时结合实体名称语义相似匹配和相同类型实体去重，从名称变体和同类型实体重叠两个方向减少重复实体累积。
  • 记忆辅助回答假设性问题：回答用户问题时，MemoryBear 现在可以围绕假设性场景进行推理，利用记忆上下文生成基于已存信息的合理回答。
  • 首页记忆标签：首页新增记忆标签入口，为用户提供当前工作区活跃记忆簇的快速可视化索引。

3. 知识库 📚

  • MCP 服务 —— CSV 导入内容完整性修复：修复 MCP 服务 CSV 导入路径中部分行未写入知识库的缺陷，导入现在正确摄入完整数据集。
  • 飞书知识库 —— 404 文件预览/下载：修复飞书知识库中返回 404 的文件阻断预览与下载流程的问题，错误现在得到优雅处理并提供适当的用户提示。
  • 分块大小范围扩展至 6–12000 字符（父子分块策略）：使用父子分块策略的知识库文档允许分块大小上限扩展至 12000 字符，为长文档检索提供更高保真度的上下文窗口。
  • 混合检索与图谱检索 —— Rerank 错误修复：修复混合检索与图谱检索时出现的运行时错误（Failed to rerank documents: 'detail'），该错误导致重排序未能完成，影响结果质量。
  • 文档解析中删除 —— 任务释放修复：修复关键任务管理缺陷：文档在解析过程中被删除后，解析任务仍滞留在 document 队列中无限运行；解析任务现在能检测删除信号并释放占用的槽位，确保队列中的其他任务可以继续执行。
  • 召回测试 —— 混合模式未设置召回数量报错修复：修复在默认混合检索模式下运行召回测试时未显式设置召回数量会报错的缺陷；召回数量现在优雅默认，测试可正常完成。

4. 稳健性与任务管理 🔧

  • Celery 检查性任务 —— 设置触发互斥锁：Celery 检查性任务在触发时现在设置互斥锁，防止并发重复执行在同一检查窗口中竞争，避免产生不一致状态。

🧭 未来展望

MemoryBear v0.3.6 社区版标志着平台三大子系统——工作流、记忆与知识库——不再各自演进，而是开始相互强化。工作流引擎不再只是自动化层：触发器节点、工具发布与缓存管理，让它正在成为其他子系统可接入的协调底座。记忆层凭借重构的反思流水线、双轨去重、实体 Embedding、时间线优化与新增的 MCP 服务端点，正从一个功能特性成长为真正的基础设施。知识库则以更宽的分块范围、修复的检索路径与更可靠的文档处理，在原型级与生产级文档智能之间的差距上迈出关键一步。

记忆 MCP 服务的引入对开源社区尤为重要。通过将记忆能力以标准 MCP 端点对外暴露，MemoryBear 现在可以参与更广泛的 AI 编排生态——任何 MCP 兼容的客户端或工具都可以直接对记忆系统进行读写。结合动态权重引擎对假设性问题的处理能力，记忆层不再是被动存储，而是多智能体架构中的主动推理参与者。下一步将推进记忆 MCP 服务上 ModelScope MCP 广场，扩大社区可见度与生态覆盖面。

Assets 2
Loading
0 Join discussion