Repository navigation
MemoryBear v0.5.4 Community Release Notes — Strengthening Memory Foundations
Codename: JingGe (惊戈 · Striking Edge)
Release Date: November 4, 2026
🧠 1. Memory Capabilities
- Eight New Ontology Dimensions for Permanent Memory: Added eight information dimensions, including educational background, professional background, health background, interpersonal relationships, group relationships, important-object relationships, pet relationships, and special scenarios. This expands the system's understanding of users from isolated facts into a more complete network of personal backgrounds and relationships.
- Unified View of Default and Custom Ontologies: Standardized the ontology engineering interface into a unified "Default Ontology + Custom Ontology" view, allowing built-in user profile structures and business-specific extension fields to be managed and compared in one place.
- Independent RAG Memory Spaces: Added support for creating independent memory spaces. Memory writing, retrieval, statistics, and management are isolated across spaces, with additional support for separating memory data by end user.
- Optimized Memory Writing Workflow: Unified asynchronous memory-writing workflows across different entry points, with support for task status tracking, failure retries, and duplicate task deduplication.
- Improved Memory Community Graph Visualization: Integrated the new community visualization capabilities to improve the visual presentation of memory community structures.
🔌 2. Agent Skills & Sandbox
- Skill Management and Invocation: Added support for importing standard Skill files and managing them centrally through the Skill Resource Center. Skills can be configured and invoked in single Agents, sub-Agents within multi-Agent systems, and Agent nodes in workflows. Skill scripts can run in isolated sandbox environments.
- Dedicated DeepSeek Harness (DSH) Plugin Support: MemoryBear now offers a dedicated integration plugin developed specifically for DeepSeek Harness. This plugin further expands the platform's integration capabilities, allowing users to seamlessly connect to and use related features.
- Independent Sandbox Execution Service: Upgraded sandbox execution into an independent service and integrated it with the Agent code interpreter and terminal tools. Supports executing Python code and terminal commands in isolated environments, maintaining runtime state within sessions, and generating and downloading file artifacts.
- Memory Toggle for Workflow Agent Nodes: Added a memory (context) toggle when associating an Agent or Agent cluster with a workflow, allowing users to enable or disable memory context as needed.
📚 3. Knowledge Base Capabilities
- Improved Document Enable/Disable Management: Disabled documents remain visible in the management list and can be re-enabled at any time. Document enablement status is now independent of Wiki processing status.
- Improved Knowledge Base Enable/Disable Management: Retrieval configurations remain accessible when a knowledge base is disabled, while actual retrieval automatically excludes the disabled knowledge base.
- Adjusted Default Configuration for API-Created Knowledge Bases: When creating knowledge bases through V1 or SSO APIs without specifying parsing configurations, scenario names and graph entity types are no longer automatically populated with preset values. Other default parsing configurations remain unchanged.
- Expanded QA Assistant Reference Materials: Added technical documentation to the MemoryBear QA Assistant's knowledge base, expanding the range of materials available for product-related questions.
🔧 4. Platform Stability & Performance
- More Stable Background Tasks in Large-Scale Spaces: Optimized user data retrieval in large-scale user spaces, reducing memory consumption during batch processing and improving the stability of background scheduled tasks.
- Optimized Object Storage Concurrency Control: Improved connection pooling and request concurrency control for OSS, S3, and MinIO, with configurable resource limits to accommodate different deployment requirements. Introduced bounded concurrency for batch image downloads in knowledge bases and improved resource cleanup when downloads complete, are interrupted, or are cancelled.
- More Reliable CSV and QA Imports: Optimized file-reading methods in import tasks to improve the reliability of CSV and QA data imports.
- Improved Gateway API Key Validation and Management: Added uniqueness validation for Gateway API Keys to prevent duplicate entries and improve key management accuracy.
- Faster Log Retrieval: Reduced retrieval latency in large-scale log scenarios, improving query response speed and search efficiency.
🧭 5. Looking Ahead
MemoryBear will continue evolving from a memory system into a more comprehensive memory-augmented Agent infrastructure. In the next phase, AI Personality will become a key area of development. Building on ontology-driven memory, we will progressively advance capabilities for personality extraction, personality anchoring, and personality presentation and management.
Personality and cognition will remain core priorities. The Personality Anchoring Engine will continue to evolve, while Ontology Engineering will be further optimized to improve the generation of candidate user-profile fields through reasoning. Capabilities such as explicit memory, episodic memory (long-term consolidation and multimodal episodic reconstruction), predictive inference graphs, and dynamic confidence calibration will further deepen MemoryBear's understanding of users.
For Agents and knowledge bases, we plan to introduce a Skill Marketplace, upgrade Memory Skills, and explore Skill-based self-evolution and candidate management. Knowledge bases will be extended to support audio and video content, hybrid retrieval for FAQ recall, Wiki candidate topic generation, richer entity descriptions, and navigation to related files. Phase II of RAG Memory Spaces, memory activity visualization across core engines, and continued service modularization will provide stronger foundations for larger-scale deployments.
Specific capabilities and delivery timelines are subject to subsequent official release announcements.
MemoryBear v0.5.4 社区版更新公告 —— 夯实记忆基础
版本代号: 惊戈 · Striking Edge
发布日期: 2026年11月4日
🧠 一、记忆能力
- DeepSeek Harness (DSH) 专属记忆插件支持: MemoryBear 现已推出由我们专门为 DeepSeek Harness 开发的集成插件。该插件进一步扩展了平台的集成能力,用户现可通过此插件无缝接入并使用支持用户画像加载、记忆检索和存储等功能。
- 永久记忆新增八类本体维度: 新增教育背景、职业背景、健康背景、人际关系、团体关系、重要对象关系、宠物关系及特殊场景等信息维度,让系统对用户的了解从零散事实扩展为更完整的背景与关系网络。
- 「默认本体 + 自定义本体」统一视图: 本体工程界面字段统一为「默认本体 + 自定义本体」的展示形式,内置画像结构与业务扩展字段可在同一视图中对照管理。
- 新增独立 RAG 记忆空间: 支持创建独立空间,各空间的记忆写入、检索、统计和管理相互隔离,同时支持按终端用户隔离记忆数据。
- 优化记忆写入流程: 统一不同入口的异步写入流程,支持任务状态追踪、失败重试和重复任务去重。
- 优化记忆社区图谱展示: 接入新版社区展示能力,优化记忆社区结构的可视化效果。
🔌 二、Agent 技能与沙箱
- Skill 技能管理与调用: 支持导入标准 Skill 文件,并通过 Skill 资源中心统一管理;可在单 Agent、多 Agent 子 Agent 及工作流 Agent 节点中配置和调用 Skill,Skill 脚本可在独立沙箱环境中执行。
- 独立沙箱执行服务: 沙箱执行能力升级为独立服务,接入 Agent 代码解释器与终端工具,支持在隔离环境中执行 Python 代码和终端命令、保持会话内运行状态,并支持文件产物生成与下载。
- 工作流 Agent 节点记忆开关: 工作流关联 Agent 或 Agent 集群时新增记忆(上下文)开关,可按需决定是否启用记忆上下文。
📚 三、知识库能力
- 优化文档启停管理: 停用的文档仍保留在管理列表中,可随时重新启用;文档启停状态与 Wiki 处理状态相互独立。
- 优化知识库启停管理: 知识库停用期间仍可查看召回配置,但实际检索时会自动排除该知识库。
- 调整接口创建知识库的默认配置: 通过 V1 或 SSO 接口创建知识库时,如未传入解析配置,场景名称和图谱实体类型不再自动填充预设值,其他默认解析配置保持不变。
- CSV 与 QA 导入更可靠: 优化导入任务的文件读取方式,提升 CSV 与 QA 导入的稳定性。
- 扩充 QA 助手问答资料: 将技术文档纳入记忆熊 QA 助手的知识库,扩大产品问答可参考的资料范围。
🔧 四、平台稳定性与性能
- 大规模空间后台任务更稳定: 优化大规模用户空间下的用户数据读取方式,降低批量处理时的内存占用,提升后台定时任务的运行稳定性。
- 对象存储并发控制优化: 优化 OSS、S3 和 MinIO 的连接池与请求并发控制,支持按部署需要配置资源上限;知识库批量图片下载采用有界并发,并完善下载结束、中断及取消时的资源释放。
- 网关 API Key 校验与管理优化: 新增唯一标识校验,避免重复添加网关 API Key,提升密钥管理的准确性。
- 日志检索提速: 降低大规模日志场景下的检索延迟,提升日志查询的响应速度与检索效率。
🧭 五、未来展望
MemoryBear 将持续从记忆系统向更完善的记忆增强型 Agent 基础设施演进。下一阶段,AI 人格将成为重点建设方向,在本体驱动的记忆基础上,逐步推进人格提取、人格锚定以及人格展示与管理等能力。
人格与认知仍是我们的重点方向:人格锚定引擎将持续演进,本体工程将持续优化推理生成候选画像字段的效果;显性记忆、情景记忆(长期固化与多模态情景重构)、预测推演图谱和置信度动态校准等能力,也将进一步加深 MemoryBear 对用户的理解。
在 Agent 与知识库方面,我们计划推出 Skill 市场、升级 Memory Skill,并引入基于 Skill 的自我进化与候选管理;知识库将支持音频与视频内容、混合检索 FAQ 召回、Wiki 候选主题生成,以及更丰富的实体描述与相关文件跳转。RAG 记忆空间二期、覆盖各核心引擎的记忆活动展示与持续推进的服务模块化,将为更大规模的部署提供支撑。
具体能力及交付时间以后续正式发布信息为准。