Release Date: July 29, 2026 | Codename: LingXiao (凌霄 · Soaring Beyond the Clouds)
MemoryBear v0.3.14 makes memory dramatically faster, broader in reach, and easier to manage. This release centers on two major breakthroughs: the Fast Dialogue memory-write pipeline brings write latency to the millisecond level, with performance improving by more than 100× in relevant test scenarios; meanwhile, the memory retrieval engine can retrieve memories for multiple end users in a single request, extending retrieval from individual lookups to efficient batch queries. The release also expands multimodal retrieval, deploys a self-trained BERT model for emotion recognition, and introduces an isolated Agent runtime sandbox for stronger fault isolation, resource governance, and concurrency isolation. Knowledge base retrieval is now faster and more scalable, while lightweight entity relations and an evidence graph make results more connected and traceable.
🧠 I. Memory Upgrades
1. Two Core Breakthroughs: Millisecond-Level Writes, Batch Retrieval 🚀
- Memory Extraction Engine Upgrade — Fast Memory-Write Pipeline: Fast Dialogue adds conversation overwrite and cleanup mechanisms and integrates retrieval support, giving applications more complete control over rapidly changing dialogue context. The optimized pipeline brings memory-write latency to the millisecond level, with measured performance improving by more than 100× in relevant rapid-write tests.
- Memory Retrieval Engine Upgrade: The memory retrieval API now accepts multiple
end_user_idvalues in a single request, eliminating repeated calls when retrieving memory across users and establishing a foundation for cross-user memory aggregation and analysis. Batch queries use a new aggregated response structure, while the existing single-user response format remains unchanged for backward compatibility.
2. Memory Management & Visibility
- Forgetting Engine Upgrade: MemoryBear now notifies users when memory storage is approaching its limit, helping teams act before capacity becomes a constraint. The redesigned forgotten-memory page also provides a clearer, more efficient cleanup and maintenance experience.
- Memory Tag Display: Memory pages now show the tags associated with each memory, making different types of content easier to recognize, distinguish, and manage.
3. Multimodal & Recent Memory Retrieval
- More Timely Short-Term Memory: The short-term memory write path has been optimized so recent conversations become available for retrieval sooner, improving continuity in fast-moving interactions.
- Multimodal & Source-Content Retrieval: Perceptual memory can now return original images and understand their content. Pruned memories can also trace back to the original text, providing richer evidence and context for retrieved results.
- Recent-Conversation Retrieval Optimization: Retrieval now handles memories from recent conversations more efficiently, improving both query speed and relevance.
4. Memory Intelligence & Reliability ✨
- Self-Trained BERT Emotion Recognition: MemoryBear now deploys and integrates a self-trained BERT model to identify emotional signals in content, laying the foundation for more natural emotional responses and personalized interactions.
- More Reliable Reflection Retries: Reflection retry processing now carries more complete user context, improving task traceability and recovery reliability.
⚙️ II. Application & Platform Upgrades
- Service Concurrency Optimization: Core interfaces now use a more consistent asynchronous execution model, reducing blocking operations and improving responsiveness under concurrent workloads.
- Faster Access to Frequently Used Configurations: Workflow execution snapshots and workspace model options are now reused more efficiently, reducing repeated data queries and accelerating common operations.
- Faster, Safer Code Execution: The code execution sandbox delivers better runtime performance and now supports non-privileged operation, strengthening security without sacrificing usability.
- Isolated Agent Runtime Sandbox: A standalone Agent runtime sandbox moves the Agent main loop into an isolated environment, separating Agent execution failures from the core service and enabling finer-grained resource governance and concurrency isolation. Tool requests are forwarded through a proxy while business data remains persisted in the core service; sandbox pre-warming and scheduling improvements further shorten startup waits and strengthen execution reliability.
- Expanded Model Marketplace: New models broaden support for multimodal understanding, ultra-long context, and complex reasoning, giving application builders more options for different workloads.
📚 III. Knowledge Base Upgrades
- Faster, More Scalable Knowledge Retrieval: The knowledge retrieval pipeline now processes multiple retrieval stages concurrently, improving response speed and throughput for complex queries and multi-knowledge-base scenarios. Existing capabilities, including GraphRAG, remain compatible throughout the upgrade.
- Lightweight Entity Relations & Evidence Graph: A new lightweight evidence graph helps MemoryBear understand relationships across knowledge content and incorporate relevant connections into retrieval when needed. Results remain traceable to their original source while providing richer relational context for more grounded, explainable answers.
🧭 Looking Ahead
v0.3.14 marks two important shifts in the speed and reach of memory. The Fast Dialogue pipeline brings memory-write latency to the millisecond level, turning rapid updates into a near-real-time capability. At the same time, the memory retrieval engine moves beyond individual lookups to efficient batch retrieval across multiple end users, enabling more effective memory aggregation and analysis. Together with capacity alerts, source-content retrieval, and stronger recent-conversation recall, these advances make memory faster, broader, and easier to control.
Next, we will introduce the first phase of emotional response strategies and a memory-activity experience that makes recent writes and reads easier to follow. Agents and workflows will also provide clearer recall details, helping developers understand how relevant memories are found and used.
We will continue deepening cross-modal association and dynamic memory-value weighting, while expanding Skill integration, multi-user memory processing, and write performance. These efforts will make memory more explainable, adaptive, and ready for a broader range of intelligent applications.
MemoryBear v0.3.14 社区版 发布说明 —— 凌霄拓境
发布日期： 2026年7月29日 ｜ 版本代号： 凌霄（LingXiao · Soaring Beyond the Clouds）
MemoryBear v0.3.14 让记忆写得更快、检索范围更广，也更易管理。本版本围绕两项核心突破展开：记忆快速写入流水线将写入时延降至毫秒级，相关测试场景下性能提升超过 100 倍；记忆检索引擎则支持在一次请求中查询多个终端用户的相关记忆，为跨用户记忆汇总和综合分析提供基础能力。同时，多模态检索进一步增强，自训练 BERT 模型带来内容情绪识别；新增独立 Agent 运行沙箱，进一步增强故障隔离、资源治理和并发隔离。知识库检索性能与并发能力进一步提升，轻量实体关系和证据图谱则让检索结果更具关联性与可追溯性。
🧠 一、记忆能力升级
1. 双核心突破：毫秒级写入，批量检索 🚀
- 记忆萃取引擎升级 - 记忆快速写入流水线：Fast Dialogue 新增对话覆盖与清理机制并接入检索能力，让应用能够更完整地管理快速变化的对话上下文。优化后的快速写入流水线将记忆写入时延降至毫秒级，相关快速写入测试场景下性能提升超过 100 倍，实现近实时记忆更新。
- 记忆检索引擎升级：记忆检索 API 现支持在一次请求中传入多个终端用户标识（
end_user_id），单次请求即可批量获取多个终端用户的相关记忆，显著减少多用户场景下的重复接口调用，并为跨用户记忆汇总和综合分析提供基础能力。批量查询采用新的聚合响应结构，原有单用户查询的返回格式保持不变，确保已有接入平滑兼容。
2. 记忆管理与可见性
- 遗忘引擎升级：新增记忆容量提醒机制，当记忆存储接近上限时主动发出提示，帮助团队提前维护容量；遗忘记忆页面同步完成交互升级，让记忆清理和存储空间管理更加清晰高效。
- 用户记忆标签展示：记忆页面新增标签信息展示，帮助用户更快识别、区分和管理不同类型的记忆内容。
3. 多模态与近期记忆检索
- 短期记忆时效提升：优化短期记忆写入链路，缩短近期对话进入可检索状态的时间，让连续交互中的上下文衔接更加及时。
- 多模态与原始内容检索：感知记忆现在可以返回原始图片并理解其中的内容；剪枝记忆也能够回溯原始文本，为检索结果提供更充分的依据和上下文。
- 近期会话检索优化：优化近期会话相关记忆的查询与匹配过程，进一步提升检索速度和内容相关性。
4. 记忆智能与可靠性 ✨
- 部署自训练 BERT 情绪识别模型：完成自训练 BERT 模型的部署与接入，支持识别内容中的情绪信息，为后续更自然的情绪回应与个性化交互奠定基础。
- 反思重试可靠性提升：反思重试过程现在具备更完整的用户上下文，提升任务追踪和异常恢复的可靠性。
⚙️ 二、应用与平台升级
- 服务并发能力优化：核心接口采用更加一致的异步执行方式，减少阻塞式调用，提升并发场景下的响应能力。
- 常用配置访问提速：工作流执行配置和工作区模型选项能够得到更高效的复用，减少重复数据查询，加快常用操作响应速度。
- 代码执行沙箱优化：代码执行沙箱在提升运行性能的同时支持非特权模式运行，兼顾使用效率与安全性。
- 独立 Agent 运行沙箱：新增独立 Agent 运行沙箱模块，将 Agent 主循环迁移至独立环境执行，实现 Agent 运行与主服务的故障隔离，支持更精细的资源管控与并发隔离；工具请求通过代理转发，业务数据仍在主服务持久化存储。本版本进一步通过沙箱预热与运行调度优化缩短启动等待时间，增强运行可靠性。
- 模型广场扩展：新增多款模型，进一步覆盖多模态理解、超长上下文和复杂推理场景，为不同应用负载提供更丰富的选择。
📚 三、知识库能力升级
- 知识库检索性能升级：全面优化知识库召回链路，支持多个检索环节并行处理，显著提升复杂查询和多知识库场景下的响应速度与并发承载能力。GraphRAG 等现有检索能力保持兼容，升级过程不会影响已有知识库的正常使用。
- 轻量实体关系与证据图谱：新增轻量实体关系与证据图谱能力，帮助系统理解知识内容之间的关联，并在检索时按需融合关系信息。检索结果可追溯至原始内容，同时提供更丰富的关联上下文，让回答更有依据、更易理解。
🧭 未来展望
v0.3.14 同时实现了记忆速度与检索范围的两次重要跃升。记忆快速写入流水线将写入时延降至毫秒级，让记忆更新迈向近实时；记忆检索引擎则从单用户查询扩展到多终端用户批量检索，为跨用户记忆汇总与分析提供支撑。配合遗忘引擎升级、原始内容回溯和近期会话检索优化，记忆变得更快、覆盖更广，也更易掌控。
接下来，我们将落地第一阶段情绪回应策略，并推出记忆活动体验，让近期的记忆写入与读取过程更加清晰。Agent 和工作流也将逐步展示召回详情，帮助开发者理解相关记忆如何被找到和使用。
我们还将持续深化跨模态记忆关联与动态价值权重能力，并扩展 Skill 集成、多用户记忆处理和写入性能，让记忆变得更加可解释、更具适应性，也能服务更广泛的智能应用。