MemoryBear v0.2.6 Release Notes — Sharpened Edge, Broader Reach

[keeees]

MemoryBear v0.2.6 Release Notes — Sharpened Edge, Broader Reach

Release Date: March 7, 2026 | Codename: TingJian (听剑 · Listening to the Sword)

MemoryBear v0.2.6 builds upon the foundation of v0.2.5 with a sweeping set of enhancements across application capabilities, memory intelligence, and platform resilience. This release introduces workflow import adaptation for Dify, multimodal input/output support including voice and file types, and a new memory pruning module — significantly expanding the system's versatility. With over a dozen feature additions and critical bug fixes spanning memory, applications, and knowledge base modules, v0.2.6 represents a major step toward production-grade robustness and richer interaction paradigms.

---

🚀 I. Core Upgrade Overview

1. Workflow & Application Framework

• Workflow Import Adaptation (Dify): Added compatibility layer for importing Dify workflow definitions, enabling seamless migration of existing workflow configurations into MemoryBear's orchestration engine.
• Field Character Limits & Validation Rules: Introduced configurable character length limits on workflow fields, with validation rules defined by product specifications to ensure data integrity at the input layer.
• Application Cloning (Agent, Workflow, Cluster): Users can now duplicate Agent, Workflow, and Cluster applications with a single action, preserving all configurations and accelerating iterative development.
• Conversation Variables (Debug + Share): Added support for conversation-scoped variables in both debug and share modes for Agent and Workflow applications, enabling stateful multi-turn interactions with persistent context.
• Streaming message_id in Chat API: The streaming output of the Chat API now includes message_id in each response chunk, allowing clients to reliably track and correlate messages in real-time conversations.

---

2. Multimodal & Interaction 💬

• Voice Input & Output: Applications now support voice as both an input and output modality, opening the door to conversational AI experiences beyond text.
• File Type Input Support (Voice, File, Video): The application input layer has been extended to accept voice recordings, generic files, and video uploads, enabling richer multimodal interaction scenarios.

---

3. Model & Intelligence 🧠

• Model Vision & Omni Differentiation: The model configuration system now distinguishes between vision-capable models and Omni models, ensuring correct capability routing and preventing mismatched model-task assignments.
• Education Memory & Companion Toy Scene Presets: Added ontology configuration presets for education memory and companion toy scenarios, providing out-of-the-box domain-specific memory structures for these verticals.
• Ontology Default Identifier: Ontology configuration presets now support a default identifier flag, making it easier to manage and select baseline configurations across scenes.
• Memory Configuration Default Identifier: Memory configurations now carry a default flag, allowing the system to automatically apply baseline settings when no custom configuration is specified.

---

4. Memory Intelligence 🔬

• Memory Pruning Module: Introduced a dedicated memory pruning engine that intelligently trims redundant or low-value memory entries, optimizing storage efficiency and retrieval relevance over time.
• RAG Quick Retrieval with Memory (Deep Think & Normal Reply): The RAG workspace now supports quick retrieval in both deep thinking and normal reply modes, with integrated memory functionality under the memory verification module for enhanced contextual accuracy.

---

5. Robustness & Bug Fixes 🔧

Model Management

• Custom Model API Key Configuration: Fixed an issue where adding a custom model under a new provider prevented batch API key configuration in the model list, with error "Provider dashscope model configuration not found". Custom models can now properly inherit provider-level key management.

Knowledge Base

• Download Original Content Error: Resolved frontend API error when attempting to download original content for non-source documents (user-uploaded files) in knowledge base management.
• Share Disable Prompt Text: Updated the prompt text displayed when knowledge base sharing is disabled to provide clearer user guidance.

User Memory

• Profile Extraction Accuracy: Improved extraction accuracy for core profile fields including name, occupation, and interests in the "About Me" section. Enhanced interest distribution analysis to provide more relevant categorization.

Long-Term Memory

• Duplicate Episodic Memory Cards: Fixed issue where episodic memory cards displayed duplicate content in the explicit memory section.
• Wrong User Attribution in Episodic Memory: Resolved bug where episodic memory cards contained QA content from incorrect users. Memory attribution now correctly matches the working memory source.

Dashboard

• Knowledge Base Count Inconsistency: Fixed discrepancy where dashboard API (/api/dashboard/dashboard_data) returned 14 knowledge bases while knowledge management page API (/api/knowledges/knowledges) returned 12. Both endpoints now return consistent counts.
• Application Count Inconsistency: Corrected application count calculation that included inactive applications (is_active='f'). Dashboard API previously returned 24, now correctly returns 20 matching the application management page.
• Total Memory Capacity Calculation: Clarified data source and query methodology for total memory capacity metric to ensure accurate reporting.
• API Call Count Tracking: Improved API call count tracking with clearer data source definition and query methodology.
• Knowledge Type Distribution Total: Fixed incorrect total count in knowledge type distribution response. Clarified the source of "memory" type in the distribution breakdown.

Infrastructure

• Celery Environment Variable Fix: Corrected misconfigured Celery environment variables that could cause task queue connection failures or unexpected worker behavior in containerized deployments.
• Database Connection Pool Leak: Fixed "idle in transaction" connection leak in Celery tasks caused by improper session management. Added automatic rollback on connection checkout to prevent stale MVCC snapshots.

---

🧭 Looking Ahead

With v0.2.6, MemoryBear crosses an important threshold — the system now handles multimodal inputs, supports cross-platform workflow migration, and features intelligent memory pruning. These are not incremental patches but foundational capabilities that redefine what the platform can do. The addition of enterprise licensing and robust validation layers signals a clear move toward production-grade deployments at scale.

The introduction of voice I/O and file-type support opens entirely new interaction paradigms. Combined with the memory pruning module and enhanced RAG retrieval, MemoryBear is evolving from a memory storage system into an active cognitive assistant — one that not only remembers but also knows what to forget and how to retrieve knowledge efficiently across modalities.

In the next release, we will focus on A2A (Agent-to-Agent) protocol support for seamless multi-agent collaboration, multimodal memory capabilities that extend memory extraction beyond text into voice and visual domains, and application import/export functionality for portable deployment and configuration sharing across environments.

---

MemoryBear v0.2.6 发布说明 —— 锋芒初露，兼收并蓄

发布日期： 2026年3月7日 ｜ 版本代号： 听剑（TingJian · Listening to the Sword）

MemoryBear v0.2.6 在 v0.2.5 的基础上进行了全面升级，涵盖应用能力、记忆智能和平台稳健性三大维度。本版本引入了 Dify 工作流导入适配、多模态语音与文件输入输出支持，以及全新的记忆剪枝模块——显著拓展了系统的应用边界。通过十余项功能新增和覆盖记忆、应用、知识库模块的关键缺陷修复，v0.2.6 标志着 MemoryBear 向生产级稳健性和更丰富交互范式迈出的重要一步。

---

🚀 一、核心升级概览

1. 工作流与应用框架

• 工作流导入适配（Dify）：新增 Dify 工作流定义的兼容导入层，支持将已有工作流配置无缝迁移至 MemoryBear 编排引擎。
• 字段字数限制与校验规则：为工作流字段引入可配置的字符长度限制，校验规则由产品规范定义，确保输入层的数据完整性。
• 应用复制（Agent、工作流、集群）：用户可一键复制 Agent、工作流和集群应用，完整保留所有配置，加速迭代开发效率。
• 对话变量（调试+分享）：Agent 和工作流应用在调试与分享模式下均支持对话级变量，实现有状态的多轮交互与持久上下文。
• Chat 接口流式输出返回 message_id：Chat API 的流式输出现在在每个响应块中包含 message_id，便于客户端在实时对话中可靠地追踪和关联消息。

---

2. 多模态与交互 💬

• 语音输入与输出：应用现已支持语音作为输入和输出模态，为超越文本的对话式 AI 体验打开大门。
• 文件类型输入支持（语音、文件、视频）：应用输入层扩展支持语音录音、通用文件和视频上传，实现更丰富的多模态交互场景。

---

3. 模型与智能 🧠

• 模型视觉与 Omni 区分：模型配置体系现已区分视觉能力模型与 Omni 模型，确保能力路由准确，避免模型与任务的错误匹配。
• 教育记忆与陪伴玩具场景本体预设：新增教育记忆和陪伴玩具场景的本体配置预设，为这些垂直领域提供开箱即用的领域专属记忆结构。
• 本体配置默认标识：本体配置预设现支持默认标识标记，便于跨场景管理和选择基线配置。
• 记忆配置默认标识：记忆配置新增默认标记，系统可在未指定自定义配置时自动应用基线设置。

---

4. 记忆智能 🔬

• 记忆剪枝模块：引入专用记忆剪枝引擎，智能裁剪冗余或低价值记忆条目，持续优化存储效率和检索相关性。
• RAG 快速检索集成记忆（深度思考与正常回复）：RAG 空间现支持深度思考和正常回复两种模式下的快速检索，并在记忆验证模块中集成记忆功能，提升上下文准确性。

---

5. 稳健性与缺陷修复 🔧

模型管理

• 自定义模型 API Key 配置错误：修复在新增模型厂商下添加自定义模型后，无法在"模型列表"中批量配置 API Key 的问题（报错"未找到供应商 dashscope 的模型配置"）。自定义模型现可正确继承厂商级别的密钥管理。

知识库管理

• 下载原始内容接口错误：修复非源文档（用户上传文件）在知识库管理中尝试"下载原始内容"时的前端接口错误。
• 分享停用提示文案：更新知识库分享停用后的提示文案，提供更清晰的用户引导。

用户记忆

• 档案提取准确性：提升核心档案字段（姓名、职业、兴趣）在"关于我"部分的提取准确性。优化兴趣分布分析，提供更相关的分类结果。

长期记忆

• 情景记忆卡片重复：修复显性记忆中情景记忆卡片显示重复内容的问题。
• 情景记忆用户归属错误：解决情景记忆卡片中包含非该用户 QA 内容的问题。记忆归属现已正确匹配工作记忆来源。

工作空间首页

• 知识库数量不一致：修复仪表盘 API（/api/dashboard/dashboard_data）返回 14 个知识库，而知识库管理页面 API（/api/knowledges/knowledges）返回 12 个的差异。两个端点现返回一致的计数。
• 应用数量不一致：修正应用数量计算包含非活跃应用（is_active='f'）的问题。仪表盘 API 此前返回 24，现正确返回 20，与应用管理页面一致。
• 总记忆容量计算：明确总记忆容量指标的数据来源和查询方法，确保准确报告。
• API 调用次数追踪：改进 API 调用次数追踪，明确数据来源定义和查询方法。
• 知识库类型分布总数：修复知识库类型分布响应中的总数错误。明确分布细分中"memory"类型的来源。

基础设施

• Celery 环境变量修复：修正 Celery 环境变量配置错误，此前可能导致容器化部署中任务队列连接失败或 Worker 异常行为。
• 数据库连接池泄漏：修复 Celery 任务中因会话管理不当导致的"idle in transaction"连接泄漏。新增连接检出时自动回滚机制，防止陈旧的 MVCC 快照。

---

🧭 未来展望

v0.2.6 让 MemoryBear 跨越了一个重要门槛——系统现已支持多模态输入、跨平台工作流迁移和智能记忆剪枝。这些并非增量补丁，而是重新定义平台能力边界的基础性功能。企业版授权和完善的校验层的加入，清晰地表明了向规模化生产级部署迈进的方向。

语音输入输出和文件类型支持的引入开启了全新的交互范式。结合记忆剪枝模块和增强的 RAG 检索，MemoryBear 正从记忆存储系统进化为主动认知助手——不仅能记住，还懂得遗忘，并能跨模态高效检索知识。

下一版本将聚焦 A2A（Agent-to-Agent）协议支持以实现无缝多智能体协作、多模态记忆能力将记忆提取从文本扩展至语音和视觉领域，以及应用导入导出功能以支持跨环境的便携部署和配置共享。

---

This discussion was created from the release MemoryBear v0.2.6 Release Notes — Sharpened Edge, Broader Reach.