MemoryBear v0.4.4 Community Release Notes — Trace Every Path, Understand Every Memory

Release Date: September 2, 2026 | Codename: XunMai (循脉 · Tracing the Meridian)

MemoryBear v0.4.4 brings improved emotional awareness, retrieval transparency, and platform stability to the Community edition. This release introduces conversation-level emotion intelligence, end-to-end recall inspection, LiteSkills write-quota management, and an independently deployable knowledge service, while improving authentication, SSO, workflow observability, and troubleshooting.

🚀 I. Core Upgrade Overview

1. Emotion Intelligence & Recall Transparency 🧠

• Conversation-Level Emotion Intelligence: Added conversation-level emotion storage and aggregation, with multilingual presentation, data quality reports, and trend analysis for richer emotional context in long-term memory.
• Improved Question-Answer Extraction: Improved matching between multiple preceding questions and a standalone answer, reducing incorrect associations during memory extraction.
• End-to-End Recall Chain Visualization: Added complete recall-chain inspection across all retrieval modes, covering query preprocessing, keyword and semantic recall, fusion, reranking, and final ranking, with detailed scores and hit sources.

2. LiteSkills Quota Management 💫

• LiteSkills Write Quota Management: Added dedicated write-quota management, supporting registration grants, usage deduction, failed-write refunds, and quota tracking throughout the write lifecycle.

3. Knowledge Services & Platform Foundations ⚙️

• Standalone **mem-knowledge** Service: Knowledge parsing, synchronization, QA import, and Evidence Graph capabilities are now provided through an independently deployable mem-knowledge service while maintaining API compatibility. PDF parsing is also streamlined with a lightweight default parser, with MinerU and TextLn available when needed.
• Deployment & Worker Health Monitoring: Added application startup, liveness, readiness, and Celery Worker health checks to help identify unhealthy services and dependencies earlier.
• Authentication & Authorization Improvements: Unified authentication and authorization through a dedicated gateway, strengthening identity validation, rate limiting, and security controls.
• Cached API Key Account Lookups: Added caching for API Key account lookups to improve performance and reduce repeated downstream requests.

4. Workflow Execution & Exception Handling 🔄

• Workflow Execution Logs: Standard workflows now provide node execution and runtime logs, making historical runs easier to trace and troubleshoot.
• Streaming LLM Exception Branches: LLM nodes now support exception handling and exception branches during streaming responses, enabling more reliable failure handling in workflows.

5. SSO & Notification Operations 🔔

• Hardened SSO API Contracts: Improved SSO parameter validation, business validation, exception handling, error messages, and documentation for more consistent integrations.
• Audience-Aware Alert Dispatch: Improved alert distribution to deliver events to their intended audiences and reduce missed or unintended notifications.
• More Reliable SSO Tenant Provisioning: Improved tenant creation and SpeedBear Key binding with automatic retries, rollback, fallback policies, and clearer failure messages.

6. Robustness & Bug Fixes 🔧

• Agent Orchestration State Fixes: Fixed empty message_id issues during model calls and abnormal regenerate-button rendering.
• Actionable Model Invocation Errors: Model invocation failures now display specific error reasons instead of generic error messages.
• Clearer Agent Document Parsing Failures: Improved error feedback for image parsing and downloading failures during document processing.
• Stable Billing Detail Pagination: Fixed duplicate billing records caused by non-unique pagination identifiers, ensuring stable and accurate billing detail navigation.

🧭 Looking Ahead

MemoryBear v0.4.4 moves long-term memory beyond simple retrieval toward a more transparent and interpretable experience, where emotional context, recall evidence, identity ownership, and runtime state can be better understood. By making both the human signals within conversations and the system paths behind retrieval more visible, this release strengthens the foundation for trustworthy memory applications.

Looking ahead, MemoryBear will continue to advance long-term memory consolidation while further separating memory and knowledge capabilities into independently deployable and governed services. Retrieval quality will also continue to improve through better result deduplication, entity preservation, embedding and reranking configuration, and stronger data-processing paths.

MemoryBear will further expand multimodal knowledge retrieval across images, text, audio, and video, while improving shared knowledge-base QA experiences and core microservice capabilities such as service registration, configuration management, and inter-service authentication. Specific capabilities and delivery timelines will be communicated through future official release announcements.

MemoryBear v0.4.4 社区版发布说明 —— 循脉知心，万流归一

发布日期： 2026年9月2日 ｜ 版本代号： 循脉（XunMai · Tracing the Meridian）

MemoryBear v0.4.4 社区版聚焦情绪感知、召回透明度与平台稳定性升级。本版本新增对话级情绪智能、完整召回链路查看、LiteSkills 写入额度管理及可独立部署的知识服务，并进一步提升认证鉴权、SSO、工作流可观测性与问题排查能力。

🚀 一、核心升级概览

1. 情绪智能与召回透明度 🧠

• 对话级情绪智能：支持在对话层面存储和聚合情绪信号，并提供多语言展示、数据质量报告和趋势分析，让长期记忆更完整地保留交流中的情绪语境。
• 多问题与单答案提取优化：优化多个上文问题与独立答案之间的匹配逻辑，减少记忆提取过程中的错误关联。
• 完整召回链路可视化：支持查看记忆在各类检索模式下的完整召回链路，覆盖查询预处理、关键词与语义召回、融合、重排及最终排序，并展示相关分数与命中来源。

2. LiteSkills 额度管理 💫

• LiteSkills 写入额度管理：新增独立写入额度管理，支持注册赠送、使用扣减、失败回补及额度消耗追踪，完善写入额度全生命周期管理。

3. 知识服务与平台基础能力 ⚙️

• 独立 **mem-knowledge** 知识服务：解析、同步、QA 导入及 Evidence Graph 等能力现由可独立部署的 mem-knowledge 服务承载，同时保持 API 兼容。PDF 解析默认采用轻量方案，并保留 MinerU、TextLn 等解析能力按需使用。
• 应用与 Worker 健康监测：新增应用启动、存活、就绪及 Celery Worker 健康检查，帮助更早发现异常服务及依赖状态。
• 认证与鉴权能力升级：通过独立网关统一处理认证与鉴权，强化身份校验、限流及安全控制。
• API Key 账户查询缓存：为 API Key 账户查询增加缓存机制，提升查询性能并减少重复请求。

4. 工作流执行与异常处理 🔄

• 工作流运行日志：普通工作流支持查看节点执行及运行日志，便于追踪历史运行记录和排查问题。
• 流式 LLM 异常分支：LLM 节点支持流式响应场景下的异常处理与异常分支执行，提升工作流失败路径的处理能力。

5. SSO 与通知运营 🔔

• SSO 接口规范加固：完善 SSO 参数校验、业务校验、异常处理、错误提示及相关文档，提升集成一致性。
• 按受众派发告警通知：优化告警事件分发机制，支持按照目标受众进行派发，降低错发和漏发风险。
• SSO 租户创建稳定性优化：优化租户创建及 SpeedBear Key 绑定流程，支持自动重试、异常回滚、降级策略及明确的失败提示。

6. 稳定性与缺陷修复 🔧

• Agent 编排状态修复：修复模型调用 message_id 为空及重新生成按钮组显示异常的问题。
• 模型调用错误原因展示：模型调用失败时展示具体失败原因，帮助快速定位问题。
• Agent 文档解析失败反馈优化：优化图片解析及下载失败时的错误提示，便于定位文档处理问题。
• 账单详单稳定分页：修复因分页标识不唯一导致的账单记录重复问题，确保账单详单分页稳定准确。

🧭 未来展望

MemoryBear v0.4.4 推动长期记忆从“能够被召回”进一步走向“可理解、可追溯”。通过让对话中的情绪语境与记忆背后的召回路径更加透明，本版本进一步夯实了可信记忆应用的基础。

下一阶段，MemoryBear 将持续深化长期记忆固化能力，并进一步推动记忆与知识能力向可独立部署、独立治理的服务演进。同时，我们将持续提升检索质量，在结果去重、实体保留、Embedding 与 Rerank 配置以及数据处理链路等方面进行优化。

MemoryBear 还将进一步拓展图片、文本、音频和视频等多模态知识检索能力，优化共享知识库 QA 体验，并持续完善服务注册、配置管理、服务间认证等微服务基础能力。具体能力及交付时间将以后续正式发布信息为准。