Skip to content

MemoryBear v0.3.10 Community Release Notes — Clarity in Every Detail

Choose a tag to compare

View all tags
@keeees keeees released this 05 Jul 07:06
· 1025 commits to main since this release
v0.3.10
157ad9c

Release Date: July 1, 2026 | Codename: DongWei (洞微 · Insight into the Subtle)

Building on v0.3.9's push toward platform-grade robustness, v0.3.10 sharpens how MemoryBear reasons over memory. Reflection alias merging is now gated behind high-confidence LLM judgment, the once black-box reflection pipeline gains full-chain snapshots for end-to-end traceability, short-term memory adds a recent-session dimension, and retrieval and writes grow more precise and resilient.

⚠️ Breaking Changes

  • Removed endpoints: /v1/memory/write/status (async write task status) and /v1/memory/read/status (async read task status) are removed, fulfilling the deprecation announced in v0.3.9. If you still call them, migrate off these endpoints.

🚀 I. Core Upgrade Overview

1. Memory Intelligence 🧠

  • High-Confidence Alias Merging: Instead of directly folding write-stage extracted aliases onto entities, the reflection stage now asks the LLM to judge — based on the entity's information — whether an alias truly belongs to that entity, executing the merge only for high-confidence results. This dramatically improves alias-merge accuracy. A new alias-merge log surfaces the reasoning and whether each merge was correct, viewable in the reflection log on the user memory page.
  • Reflection Snapshots: Reflection tasks now capture a snapshot recording each task's recall results and the LLM's output, turning a previously black-box process into a fully transparent, end-to-end traceable pipeline that makes it easy to pinpoint weak spots and improve them.
  • Active-User Detection by Write Time: The reflection stage now determines active users by last write time rather than last conversation time, covering users from all sources — including new API and MCP users.
  • Reflection Enabled by Default: Space-level default memory configuration is optimized to make reflection more readily available out of the box.

2. Short-Term Memory & Retrieval 🔍

  • Recent-Session Retrieval Dimension: The workflow memory extraction/retrieval mode adds a "recent sessions" dimension, drawing on historical session information alongside the current conversation context to enrich short-term memory sources so replies better fit the full dialogue context.
  • Retrieval Strategy & Naming: Memory retrieval strategies are refined with upgraded mode naming and improved compatibility.
  • Unified Memory Write with Conversation Time: Memory write is unified and upgraded with stronger support for conversation timestamp information.
  • Memory MCP Persistence: The memory MCP capability gains persistence for memory content, improving traceability and stability.

3. Knowledge Base & Document Processing 📚

  • Markdown Structured Parsing & Smarter Chunking: Knowledge base text parsing and chunking are upgraded with new Markdown structured parsing and optimized multi-type chunking strategies, better preserving document structure for improved retrieval.
  • Batch Chunk Import API: A new batch chunk import endpoint supports creating general chunks and parent-child chunks in bulk simultaneously.
  • Smart Metadata Filtering & Adaptive Model Params: Knowledge retrieval is upgraded with intelligent metadata filtering and adaptive model parameters.

4. Performance & Stability ⚡

  • System Model Space Adaptation: System models can be configured in the admin backend and bound to a space's model list, with Agents and workflows able to invoke system models seamlessly.
  • Connection Pool Configuration: Connection pool parameters are now configurable and centrally managed, improving connection stability under heavy load.
  • Clustering Task Optimization: Clustering task computation is optimized on top of the new connection pool design, easing the load on compute resources.
  • Write Pipeline Resilience: The memory write pipeline's exception handling and retry mechanisms are improved, boosting the stability and success rate of memory write tasks.
  • Config Time Format Standardization: The memory configuration API standardizes its time format for improved data consistency.

5. Interaction & Experience 🎨

  • Memory Card end_user_id Display: Memory library cards now surface end_user_id, with refined user-info display logic that shows other_name and end_user_id compatibly.
  • AI Message Favorites: Message operations are enhanced with support for favoriting AI messages and a range of content management actions.
  • "Reply" Node Rename: The conversation workflow's "End" node is renamed to "Reply" for clearer node semantics and a better authoring experience.
  • Workflow Turn Replay: Conversation workflow debugging now supports repeatedly running a specified conversation turn.
  • Extraction Debug Panel Auto-Expand: The extraction engine's trial-run "Create Nodes and Edges" step gains a start event, automatically expanding the "Create Entity Relationships" panel once extraction results are identified.

🧭 Looking Ahead

v0.3.10 marks a decisive turn toward precision and transparency in how MemoryBear reasons over memory. By gating alias merges behind high-confidence LLM judgment, capturing full-chain reflection snapshots, and unifying active-user detection across every source, the reflection pipeline moves from a black box to an auditable, tunable system — a hallmark of platform-grade maturity.

The forward-looking thread of this release is context depth and control. The new recent-session retrieval dimension, unified memory writes carrying conversation time, and a hardened write pipeline lay the groundwork for deeper short-term and working memory that spans multiple conversations, along with early preparation for richer cross-modal association.

Next, we will deepen knowledge processing with broader document format support — including PDF and Word — cross-knowledge-base recall, and customizable metadata for ontology modeling. On the platform side, expect continued performance and stability gains and a more consistent API experience.

MemoryBear v0.3.10 社区版 发布说明 —— 明察秋毫

发布日期： 2026年7月1日 ｜ 版本代号： 洞微（DongWei · Insight into the Subtle）

在 v0.3.9 迈向平台级稳健的基础上，v0.3.10 进一步锤炼记忆熊对记忆的推理能力。反思别名合并引入高置信度 LLM 判定，原本黑箱的反思流程新增全链路快照、实现端到端可溯，短期记忆新增「近期会话」维度，检索与写入更精准稳健。

⚠️ 不兼容变更

  • 接口移除：移除 /v1/memory/write/status（异步写任务状态查询）与 /v1/memory/read/status（异步读任务状态查询），兑现 v0.3.9 的废弃预告。如仍在调用，请及时迁移。

🚀 一、核心升级概览

1. 记忆智能 🧠

  • 别名合并高置信度判定：反思阶段先由 LLM 判断别名是否真正属于实体，仅对高置信度结果执行合并，大幅提升准确率，并新增别名合并日志可查看合并理由与对错。
  • 反思快照：为每个反思任务记录召回情况与 LLM 输出，让原本黑箱的反思全链路透明，便于快速定位并优化不足。
  • 活跃用户判定升级：反思阶段活跃用户判断从最近对话时间升级为最后写入时间，覆盖全部来源用户，包括新增的 API 与 MCP 用户。
  • 默认反思能力增强：优化空间级默认记忆配置，增强反思能力的默认可用性。

2. 短期记忆与检索 🔍

  • 「近期会话」检索维度：新增「近期会话」检索维度，结合当前对话上下文调取历史会话，丰富短期记忆数据源，让回复更贴合完整语境。
  • 记忆检索策略优化：升级检索模式命名与兼容能力。
  • 记忆写入统一升级：统一升级记忆写入能力，增强对对话时间信息的支持。
  • 记忆 MCP 持久化：记忆 MCP 能力升级，新增记忆内容持久化能力，提升记忆信息的可追溯性与稳定性。

3. 知识库与文档处理 📚

  • Markdown 结构化解析与智能切块：新增 Markdown 结构化解析并优化多类型分块策略，更好保留内容结构、提升检索效果。
  • 分块批量导入接口：新增分块批量导入接口，支持同时批量创建通用块与父子块。
  • 智能元数据过滤与模型自适配：知识检索能力升级，支持智能元数据过滤与模型参数自适配。

4. 性能与稳定性 ⚡

  • 系统模型空间适配：支持后台配置系统模型，空间绑定模型列表，Agent、工作流兼容调用系统模型。
  • 连接池配置：新增连接池配置参数并统一管理，提升高负载下的连接稳定性。
  • 聚类任务计算优化：基于全新连接池设计优化聚类任务内容计算，减轻计算资源负担。
  • 写入管线稳健性：优化记忆写入管线的异常处理与重试机制，提升记忆写入任务执行稳定性与成功率。
  • 配置接口时间格式标准化：记忆配置接口时间格式标准化，提升数据一致性。

5. 交互与体验 🎨

  • 记忆库卡片 end_user_id 展示：记忆库卡片新增 end_user_id 展示，优化用户信息展示逻辑，支持 other_name 与 end_user_id 兼容展示。
  • AI 消息收藏：增强消息操作能力，支持 AI 消息收藏与多种内容管理操作。
  • 「回复」节点更名：优化对话工作流节点命名，将「结束」节点调整为「回复」节点，提升节点语义清晰度与使用体验。
  • 对话工作流轮次重跑：对话工作流调试能力增强，支持指定轮次对话重复运行。
  • 萃取调试面板自动展开：萃取引擎配置优化，为调试试运行的「创建节点和边」新增开始事件，识别出提取结果后自动展开「创建实体关系」面板。

🧭 未来展望

v0.3.10 标志着记忆熊在记忆推理上向精准与透明的坚定转向。别名合并引入高置信度 LLM 判定、反思全链路快照可溯、活跃用户判定覆盖全部来源，反思流程由黑箱走向可审计、可调优的系统，尽显平台级成熟度。

本版本的前瞻主线是上下文的深度与可控。全新「近期会话」检索维度、携带对话时间的统一写入与更稳健的写入管线，为后续跨会话的短期与工作记忆打下基础，并为更丰富的多模态关联做前期准备。

接下来，我们将深化知识处理，带来更广泛的文档格式支持（如 PDF、Word）、跨知识库召回，以及面向本体建模的自定义 metadata 能力；平台侧将持续提升性能与稳定性，并带来更一致的 API 体验。

Assets 2
Loading
0 Join discussion