Skip to content

MemoryBear v0.3.9 Community Release Notes — Unstoppable Momentum

Latest
Latest

Choose a tag to compare

View all tags
@keeees keeees released this 26 Jun 02:57
v0.3.9
8c2a65d

Release Date: June 24, 2026 | Codename: PoJun (破军 · Breaking the Vanguard)

This release pushes MemoryBear from feature accumulation toward platform-grade robustness. Memory reflection becomes more reliable, knowledge processing handles richer file types, and retrieval gets faster. Together with stronger entity relationships and the memory service's debut on the ModelScope MCP marketplace, this version sharpens both the depth and the reach of memory capabilities.

⚠️ Breaking Changes

  • Removed endpoints: The synchronous-write /v1/memory/write/sync and asynchronous-read /v1/memory/read endpoints have been removed in this version. Please migrate to the current memory write/read endpoints before upgrading. If you need migration guidance, refer to the API documentation.
  • Deprecation notice (removal in next version): The /v1/write/status (async write task status query) and /v1/read/status (async read task status query) endpoints will be removed in the next version. Please check whether your integration calls these endpoints and plan your migration accordingly.

🚀 I. Core Upgrade Overview

1. Memory Intelligence 🧠

  • More Reliable Reflection: Memory reflection now executes more steadily and safely, sustaining stable performance even under heavy load so reflection results are produced consistently.
  • Event Timeline Editing: The reflection engine can now add, update, and delete events on an entity's description timeline, dynamically enriching existing events, appending new ones, and removing outdated details based on user input.
  • Multi-Relationship Entity Types: Build entity relationship networks across multiple dimensions, strengthening relationship analysis and surfacing latent associations.
  • More Accurate Memory Extraction: Improved the accuracy of statement extraction and simplified its field structure for cleaner, more consistent results.
  • Smarter Memory Compression: Refined the memory compression strategy to better preserve key information such as timestamps and important events, improving information completeness.
  • More Precise Deduplication: Memory merging is now more precise — part-whole and subclass-superclass cases are no longer incorrectly merged, while genuine duplicates are consolidated reliably.
  • Metadata Management: Memory metadata can now be correctly added, updated, and deleted, with more consistent structured management throughout the memory lifecycle.
  • Express Retrieval Mode: A new high-speed memory retrieval mode optimized for latency-sensitive scenarios, delivering faster responses for real-time interaction.
  • Trial-Run Memory Cleanup: Temporary memories generated during trial runs are now automatically cleaned up, preventing stale data accumulation and improving data quality.

2. Knowledge Base & Document Processing 📚

  • Structured Data Parsing: Structured files such as JSON, Excel, and CSV are now parsed through a dedicated path separate from general text, improving tabular data parsing and processing quality.
  • More Maintainable Chunking: The knowledge base chunking flow has been improved for greater stability and extensibility, making it easier to support more file types and chunking strategies going forward.
  • Parent-Child Chunking Recognition: The system now uniformly determines and reports parent-child chunking mode (parent_child_mode), for more consistent chunking behavior.
  • Structured Excel Parser: The new Excel parser recognizes sheets, merged cells, sub-table regions, and headers, converting Excel data into structured chunks with field semantics and provenance — improving Excel retrieval accuracy in RAG scenarios.

3. Application & Knowledge Recall ⚙️

  • Object Storage Support: Added object storage capability for file storage, improving scalability and reliability.
  • Unified Knowledge Recall: Workflow and Agent now share a unified knowledge-retrieval capability, delivering more consistent retrieval behavior across both and improving reuse.
  • Navigation Rename: Renamed "Application Management" to "Agent Management" in the in-workspace left navigation bar.

4. API & Ecosystem Integration 🔌

  • ModelScope MCP Marketplace: The memory MCP service is now listed on the ModelScope MCP marketplace, expanding ecosystem reach and making it easier to discover and use memory capabilities.
  • End-Users Query API: Added /v1/dashboard/end_users, which paginates the user list and memory counts for a workspace using its API key.
  • Streamlined API Key Onboarding: Creating an API key now returns the associated user identifiers in the same response, letting integrations complete identity mapping and memory association more quickly.

5. User Experience 🎨

  • Knowledge Base Detail UI: Refined the action bar on the knowledge base detail page, adjusting the layout of operation entries.
  • Workflow Canvas Layout: Moved the workflow orchestration action bar downward for a cleaner layout.

🧭 Looking Ahead

This release marks a turning point in MemoryBear's maturity. More reliable reflection, richer knowledge processing, and stronger storage together signal a shift from feature accumulation toward platform-grade robustness. The system is now better equipped to carry production workloads safely and steadily.

The introduction of express retrieval, multi-relationship entity modeling, and the ModelScope MCP listing point toward a broader trajectory: memory as an open, fast, and richly connected capability. By unifying knowledge recall across workflows and agents, MemoryBear is laying the groundwork for memory that is both deeper in understanding and wider in reach.

Looking ahead, we will deepen short-term and working memory by tying them more closely to context, prototype a cross-modal memory association engine, and build out the forgetting engine with finer-grained controls. On the platform side, expect continued performance improvements, broader system-model adaptation, and a more consistent API experience. We are also bringing repeatable single-turn replay to the trial-run window.

MemoryBear v0.3.9 社区版 发布说明 —— 势如破竹

发布日期： 2026年6月24日 ｜ 版本代号： 破军（PoJun · Breaking the Vanguard）

本次发布推动 MemoryBear 从功能堆叠迈向平台级稳健。记忆反思更可靠，知识处理可应对更丰富的文件类型，检索速度更快。配合更强的实体关系能力，以及记忆服务登陆 ModelScope MCP 广场，本版本同时强化了记忆能力的「深度」与「触达」。

⚠️ 不兼容变更（Breaking Changes）

  • 接口移除：本版本移除同步写 /v1/memory/write/sync 与异步读 /v1/memory/read 两个对外接口。升级前请迁移至当前的记忆写入 / 读取接口。如需迁移指引，请参考 API 文档。
  • 接口废弃预告（下个版本移除）：/v1/write/status（查询异步写入任务状态）与 /v1/read/status（查询异步读取任务状态）将在下个版本移除。请调用方检查是否调用了以上接口并提前规划迁移。

🚀 一、核心升级概览

1. 记忆智能 🧠

  • 反思更稳定可靠：记忆反思的执行更加稳健、安全，即使在高负载场景下也能稳步产出反思结果。
  • 事件时间线增删改：反思引擎现支持对实体描述时间线进行事件的增、删、改，能根据用户描述动态补充已有事件细节、添加新事件并删除不符的旧事件。
  • 用户实体多关系类型：支持从多个维度构建实体间关系网络，增强关系分析能力，挖掘潜在关联信息。
  • 记忆提取更准确：提升记忆陈述提取的准确性，并简化字段结构，使结果更干净、一致。
  • 记忆压缩更智能：优化记忆压缩策略，更好地保留原文中的时间、重要事件等关键信息，提升信息完整性。
  • 去重更精准：记忆合并更精准——部分–整体、子类–父类等情况不再被错误合并，真正的重复项则稳定归并。
  • Metadata 管理：支持记忆 metadata 的正确增删改，在记忆全生命周期中提供更一致的结构化管理。
  • 记忆极速检索模式：新增记忆极速检索模式，针对响应速度敏感的场景优化，为实时交互等场景提供更快的响应体验。
  • 试运行记忆清理：自动清理试运行过程中产生的临时记忆，避免无效数据长期留存，提升系统数据质量。

2. 知识库与文档处理 📚

  • 结构化数据解析：JSON、Excel、CSV 等结构化文件现通过独立于通用文本的专用链路解析，提升表格类数据的解析与处理质量。
  • 分块流程更易维护：知识库分块流程在稳定性与扩展性上得到提升，便于后续支持更多文件类型和分块策略。
  • 父子分块模式识别：系统现统一判定并返回父子分块模式（parent_child_mode），分块行为更一致。
  • 结构化 Excel 解析器：新增结构化 Excel 解析器，可识别 Sheet、合并单元格、子表区域及表头信息，将 Excel 数据转换为具备字段语义和溯源信息的结构化 Chunk，提升 RAG 场景下 Excel 数据的检索准确性。

3. 应用与知识召回 ⚙️

  • 对象存储支持：新增文件存储的对象存储能力，提升可扩展性与可靠性。
  • 统一知识召回：工作流与 Agent 现共享统一的知识检索能力，两者检索行为更一致，复用性更强。
  • 导航菜单更名：在空间内的左侧导航栏中，将「应用管理」更名为「Agent 管理」。

4. API 与生态集成 🔌

  • ModelScope MCP 广场：记忆 MCP 服务已上架 ModelScope MCP 广场，扩展生态接入能力，方便用户发现和使用记忆能力。
  • End-Users 查询接口：新增 /v1/dashboard/end_users 接口，使用工作空间 API Key 分页查询该工作空间下的用户列表与记忆数量。
  • API Key 接入更顺畅：创建 API Key 时将在同一响应中返回关联用户标识，帮助集成方更快完成身份映射与记忆数据关联。

5. 用户体验 🎨

  • 知识库详情页 UI：优化知识库详情页操作栏，调整操作入口布局。
  • 工作流画布布局：优化工作流编排页布局，将操作栏下移。

🧭 未来展望

本次发布是 MemoryBear 走向成熟的重要节点。更可靠的反思、更丰富的知识处理，以及更稳固的存储，共同标志着产品从功能堆叠迈向平台级稳健。系统已能更安全、更稳定地承载生产负载。

极速检索、实体多关系建模以及记忆 MCP 服务上架 ModelScope，勾勒出更宏大的方向——让记忆成为开放、快速且高度关联的能力。通过统一工作流与 Agent 的知识召回，MemoryBear 正为「理解更深、触达更广」的记忆能力夯实基础。

展望未来，我们将结合上下文深化短期与工作记忆、设计跨模态记忆关联联想引擎原型，并以更精细的控制完善遗忘引擎。平台侧将持续带来性能提升、更广泛的系统模型适配，以及更一致的 API 体验；同时，试运行窗口也将支持单轮对话的重复运行。

Assets 2
Loading
0 Join discussion