MemoryBear v0.4.2 Community Release Notes — Memory in Balance

[keeees]

Release Date: August 19, 2026 | Codename: XuanHeng (玄衡 · Mystic Balance)

MemoryBear v0.4.2 introduces MemoryLiteSkills, a standalone long-term memory application built for AI agents, while further upgrading memory value assessment, retrieval accuracy, writing stability, and memory management. The release also improves multimodal recall, knowledge chunking, and structured-data handling across application workflows.

---

🚀 I. Core Upgrade Overview

1. Memory Value & Retrieval Intelligence 🧠

• Dynamic Memory Value Weighting Engine Update: The engine now continuously evaluates memory value across multiple dimensions and intelligently tiers and ranks memories based on dynamic weights, further improving memory filtering, ranking, and retrieval.
• Enhanced User Memory Protection: Long-term memory safeguards have been refined to better recognize and protect user-related information, improving the safety and continuity of memory management.
• Safer Memory Management: Important deletion actions now require secondary confirmation, helping users avoid unintended operations and maintain greater control over their memories.
• More Accurate Instant Retrieval: Refined recall scope and matching strategies reduce irrelevant content and make instant retrieval results more focused and precise.
• Space-Level Memory Model Configuration: Memory model settings have been consolidated at the space level, removing fragmented configuration paths and simplifying administration.

---

2. Memory Reliability & Experience 🧭

• More Reliable Memory Ingestion: Memory writing and processing flows have been optimized to improve efficiency and stability across complex application scenarios.
• Reflection Task Alerts: When a reflection task requires attention, the message center notifies space administrators so the issue can be identified and addressed promptly.
• Cross-Modal Memory Recall in Agent Shared Sessions: Agent shared sessions can now display recalled images during long-term memory tool calls, making multimodal retrieval results more intuitive and easier to understand.
• Timeline Filtering by Time Type: The user memory timeline can be filtered by different time dimensions, enabling more focused review of memory records and lifecycle events.

---

3. MemoryLiteSkills: AI That Understands You Better Over Time 🔌

• Long-Term Memory Across Sessions and Clients: MemoryLiteSkills is a standalone long-term memory application for AI agents. Any MCP-compatible AI client can read and write personal long-term memory, allowing different AI tools to share context and understand the same user continuously over time.
• Visual Memory Management Portal: After adding SKILL.md to an Agent, users can register by email, claim an sk-mem- memory Key, and share long-term memory across multiple AI clients. The portal presents knowledge graphs, memory insights, interest distributions, and other analyses so users can see and manage what AI remembers.
• Three-Step Setup: Getting started only requires registering for a memory Key, configuring MCP, and beginning to use the service. With SKILL.md, AI can also guide users through one-click onboarding without requiring manual configuration authoring.
• Secure and Controllable by Design: MemoryLiteSkills uses dual authentication with a user Key and an underlying space Key to isolate data; end_user_id is injected by the server, and keys are never sent to the browser. Built-in rate limiting, data masking, and account deletion help ensure memories remain persistent, visible, and controllable.

---

4. Workflow & API Experience ⚙️

• Clearer Quota Feedback: Application conversations now provide clearer, more consistent guidance when available quota is insufficient, helping developers identify the cause and guide users toward resolution.
• Configurable Multi-File Inputs: Workflow file parameters now support a configurable upload limit of up to five files, giving application builders clearer control over multi-file input.
• Object Output Interoperability: Output nodes can now consume Object data emitted by iteration nodes and variable aggregators, enabling richer structured-data handoffs across workflows.

---

5. Multimodal Knowledge Processing 📚

• Context-Aware Structured Markdown Chunking: Automatic chunking now consistently merges context across headings, body text, tables, code blocks, quotations, and images, improving chunk completeness and downstream retrieval and answer quality.
• Combined Image-and-Text Knowledge Slices: Knowledge base slicing can now preserve images together with their related copy and surrounding context, strengthening the relevance and precision of multimodal retrieval.

---

🧭 Looking Ahead

With v0.4.2, MemoryBear establishes a more disciplined memory lifecycle in which information is not merely stored, but evaluated, protected, connected, and observed. These improvements support more reliable memory governance while preserving user control over important information.

The next stage will continue to improve memory continuity across applications and channels, creating a more connected and consistent experience as users move between different scenarios.

Upcoming releases will further enhance retrieval accuracy and scalability, expand memory lifecycle notifications and management experiences, and strengthen knowledge management and system reliability to support richer use cases at greater scale.

---

MemoryBear v0.4.2 社区版 发布说明 —— 玄衡定忆

发布日期： 2026年8月19日 ｜ 版本代号： 玄衡（XuanHeng · Mystic Balance）

MemoryBear v0.4.2 重点推出面向 AI 智能体的独立长期记忆应用 MemoryLiteSkills，并进一步升级记忆价值评估、检索准确性、写入稳定性与记忆管理体验。本版本同时增强多模态记忆召回、知识切片与工作流结构化数据处理，为社区用户构建更灵活、更稳健的记忆应用基础。

---

🚀 一、核心升级概览

1. 记忆价值与检索智能 🧠

• 记忆动态价值权重引擎更新：更新记忆动态价值权重引擎，从多个维度持续评估记忆价值，并根据动态权重进行智能分层与排行，进一步提升记忆筛选、排序和检索效果。
• 用户记忆保护增强：优化长期记忆保护策略，加强对用户相关信息的识别与保护，提升记忆管理的安全性和连续性。
• 记忆管理操作保护：重要删除操作新增二次确认，帮助用户避免误操作，让记忆管理更加安全可控。
• 极速检索准确性提升：优化记忆召回范围与匹配策略，减少无关内容干扰，让检索结果更加聚焦、准确。
• 空间级记忆模型配置：将记忆模型配置统一收归空间层级，减少分散配置入口，简化管理和维护流程。

---

2. 记忆可靠性与使用体验 🧭

• 记忆写入稳定性提升：优化记忆写入与处理流程，提升复杂应用场景下的处理效率和运行稳定性。
• 反思任务异常提醒：当反思任务出现需要关注的异常时，消息中心将及时通知空间管理员，帮助团队快速发现并处理问题。
• Agent 分享会话跨模态记忆召回展示：Agent 分享会话长期记忆工具调用过程中现可同步展示召回图片，使多模态检索结果更加直观、更易理解。
• 记忆时间线时间类型筛选：用户记忆时间线新增时间类型筛选，可按不同时间维度查看记忆记录及生命周期事件。

---

3. MemoryLiteSkills：让 AI 越用越懂你 🔌

• 跨会话、跨客户端的长期记忆：MemoryLiteSkills 是一款面向 AI 智能体的独立长期记忆应用。任何兼容 MCP 的 AI 客户端均可读取和写入个人长期记忆，让不同 AI 能够共享上下文，并在持续使用中越来越懂你。
• 可视化记忆管理门户：用户将 SKILL.md 文件添加到 Agent 后，可通过邮箱注册并领取一把 sk-mem- 记忆 Key，在多个 AI 客户端之间共享长期记忆。门户支持查看知识图谱、记忆洞察、兴趣分布等分析，帮助用户随时了解并管理 AI 记住了什么。
• 三步快速接入：接入仅需注册并领取记忆 Key、配置 MCP、开始使用三个步骤。配合 SKILL.md，还可由 AI 引导完成一键接入，无需手动编写配置。
• 安全、隔离、可控：MemoryLiteSkills 通过用户 Key 与底层空间 Key 双重认证实现数据隔离；end_user_id 由服务端强制注入，密钥不会下发至浏览器。同时内置限流、数据脱敏和账号注销等能力，真正做到：记得住、看得见、控得了。

---

4. 工作流与 API 体验 ⚙️

• 额度不足提示优化：优化应用对话中的额度状态反馈，在可用额度不足时提供更清晰、规范的提示，帮助开发者快速定位原因并引导用户完成后续处理。
• 可配置多文件输入：工作流文件参数现可配置上传数量上限，最多支持五个文件，为多文件输入提供更清晰的约束与控制。
• Object 输出数据互通：输出节点现可接收迭代节点及变量聚合器产生的 Object 数据，支持更丰富的结构化数据在工作流节点间传递。

---

5. 多模态知识处理 📚

• 结构化 Markdown 上下文分块：自动分块现可统一合并标题、正文、表格、代码、引用和图片的关联上下文，提升切片完整性以及后续检索和回答质量。
• 图片与文案组合切片：知识库切片现可保留图片及其关联文案和上下文，增强多模态内容之间的关联性，提高检索准确性。

---

🧭 未来展望

v0.4.2 让 MemoryBear 建立起更加有序的记忆生命周期：信息不再只是被保存，还能被衡量、保护、连接和观察。这些改进让记忆治理更加可靠，同时保留用户对重要信息的掌控。

下一阶段，我们将持续完善跨应用、跨渠道的用户记忆衔接能力，让用户在不同使用场景中获得更加连贯、一致的记忆体验。

后续版本还将进一步提升记忆检索的准确性与扩展能力，完善记忆生命周期中的通知与管理体验，并持续增强知识管理能力与系统稳定性，为更丰富的业务场景和更大规模的应用提供可靠支撑。

---

This discussion was created from the release MemoryBear v0.4.2 Community Release Notes — Memory in Balance.