Skip to content

MemoryBear v0.3.12 Community Release Notes — Reflections That Endure

Choose a tag to compare

View all tags
@keeees keeees released this 17 Jul 12:06
· 595 commits to main since this release
v0.3.12
cc05e1c

Release Date: July 16, 2026 | Codename: MingJing (明镜 · The Bright Mirror)

MemoryBear v0.3.12 sharpens the reflection engine and expands memory intelligence. Building on the resilience foundation of v0.3.11, this release strengthens reflection for heavy-load users, introduces Express Retrieval and Persona Memory as first-class Agent tools, and refines the forgetting engine with real activity metrics. On the knowledge and application side, PPT parsing upgrades to MinerU V3, a native PostgreSQL query tool joins the built-in toolkit, and the new License Management Platform lands to cover the full lifecycle of authorization.

🚀 I. Core Upgrade Overview

1. Memory Intelligence 🧠

  • Reflection Robustness for Heavy-Load Users: Reflection now tracks heavy-volume and failed users and dispatches targeted retries with multi-round redispatch, completing on time even under skewed data.
  • Agent Memory Tool Calling — Express Retrieval & Persona Memory: Agents can now natively call Express Retrieval and Persona Memory as first-class tools, cutting tool-calling latency and delivering more precise, personalized context.
  • Unified LLM Client Across Reflection Pipeline: Memory write, reflection, and query now share one LLM client, keeping model access consistent and stable under failure or config drift.
  • Attachment Nodes for Long-List Messages: Long-list or catalog-style messages that need pruning now generate independent attachment nodes linked to user and document entities, improving retrieval accuracy.
  • Pruning Deduplication via Memory Message Layer: Pruning results are persisted at the memory-message layer so identical content is never pruned twice, saving LLM calls and tokens on repeated ingestion.

2. Memory Lifecycle & Forgetting Engine 🕰️

  • Memory Node Storage Ceiling with Auto-Forget: A configurable storage cap automatically triggers the forgetting strategy once exceeded, keeping memory volume in check and preventing retrieval degradation over time.
  • Direct Memory Node Deletion from Frontend: Memory nodes can now be deleted directly from the management page, making manual maintenance and audit-driven cleanup far more efficient.
  • Activity-Aware Forgetting: The forgetting engine now knows when each memory was created and how often it's been used, so it can quietly clear away rarely-touched memories and keep the ones that matter.
  • Null-Tolerant Pruning + Default-On Migration: Pruning now tolerates null QA inputs and is enabled by default via migration — MCP, API, and custom workflow inputs all get focused, token-efficient extraction with no manual setup.

3. Knowledge Base 📚

  • QA Export to CSV: The document list page adds an "Export QA" button, enabling full-scale CSV export for batch backup, offline editing, and cross-system migration.
  • PPT Parsing Upgrade with MinerU V3: The PPT/PPTX pipeline now runs on MinerU V3 with PDF-fallback progress callbacks and stronger format-aware fault tolerance.
  • Single-KB Retrieval Rerank & Filter Optimization: Hybrid retrieval drops the redundant double-rerank and supports custom rerank models; tokenized retrieval removes the global threshold filter — balancing speed with recall completeness.

4. Applications & Workflow ⚙️

  • License Management Platform: A dedicated License Management Platform is now live, covering the full lifecycle of permissions, customer records, and authorization end-to-end.
  • Auto-Binding Memory Config on Workflow/Agent Migration: After cross-space migration, "Memory Extraction" and "Memory Storage" nodes auto-bind to the target space's default configuration — no manual reconfiguration required.
  • Native PostgreSQL Query Tool: A built-in database query tool natively supports PostgreSQL, callable directly by workflows and agents with no custom development.
  • Self-Hosted MinerU Support: Workflows and agents can now call privately deployed MinerU services, enabling structured document parsing entirely within on-prem environments.
  • HTTP Node "Import cURL": The workflow HTTP request node adds an "Import cURL" action, translating a cURL command into a fully configured request in one step.
  • SSO Workspace Default Model Fallback: SSO-created tenant workspaces without a model parameter now fall back to the global default model, ensuring memory initialization always succeeds.

5. Performance & Stability ⚡

  • Chat Full-Chain Latency Optimization: Node-level DB connection logic is refactored to cut connection overhead, shortening end-to-end response and noticeably reducing stalls in high-frequency chat and memory queries.

🧭 Looking Ahead

With v0.3.12, MemoryBear enters a phase where memory is no longer only stored and retrieved — it is watched, weighed, and gracefully retired. The reflection engine now understands its own edge cases; the forgetting engine reasons over real activity signals; and memory intelligence becomes visible to the human eye through the value-weight frontend. These changes together move the platform from feature completeness toward a more self-aware, self-healing memory substrate — one that can serve production workloads without constant hand-holding.

Meanwhile, the arrival of Express Retrieval and Persona Memory as Agent tools, alongside the native PostgreSQL query tool and self-hosted MinerU support, marks a decisive step toward MemoryBear becoming an infrastructure layer for context-aware agents — not merely a memory service. The License Management Platform completes the enterprise loop, bringing authorization, customer, and lifecycle management under one roof. The product is trending toward a platform where memory, knowledge, and workflow orchestration converge with production-grade governance.

For the next release, we will keep polishing a few key directions. Open APIs will gain finer-grained usage and rate management, so teams of every size can adopt MemoryBear with confidence. Memory growth and forgetting will be woven into a single traceable timeline, where every birth, evolution, and fading of a memory is clearly visible. Short-term recall will be sharpened further so the current conversation is understood faster and cited more accurately. On top of that, we will adopt an industry-recognized long-term memory benchmark to hold our recall quality to a higher bar, and continue refining knowledge base performance under high concurrency and pure-image scenarios. On the experience side, workflows and dialogue will support one-question-many-answers streaming responses, and parameter configurations will behave more consistently across self-run, shared, and API-triggered scenarios — closer to how people actually use them. The mirror is polished — next, it learns to see clearer, respond faster, and forget wiser.

MemoryBear v0.3.12 社区版 发布说明 —— 镜照初心，忆久弥新

发布日期： 2026年7月16日 ｜ 版本代号： 明镜（MingJing · The Bright Mirror）

MemoryBear v0.3.12 让反思更稳、让记忆更懂节制。在 v0.3.11 稳健性积累的基础上，本次版本重点补齐反思阶段对大数据量用户的处理与重试机制，把极速检索（Express Retrieval）与用户画像（Persona Memory）作为一等公民接入 Agent 工具链，并以真实活跃度指标进一步细化遗忘引擎的判定逻辑。知识库侧完成 PPT 解析对接 MinerU V3、单库检索重排与过滤逻辑优化；应用侧上线 License 管理平台、内置 PostgreSQL 查询工具、支持私有化 MinerU，并为工作流 HTTP 请求节点加入"导入 cURL"能力。

🚀 一、核心升级概览

1. 记忆智能 🧠

  • 反思引擎大数据量用户稳健性：反思流水线持续记录高数据量与失败用户，触发定向重试与多次重派，倾斜数据下也能按时完成。
  • Agent 记忆工具调用——极速检索与用户画像：Agent 原生支持极速检索（Express Retrieval）与用户画像（Persona Memory）作为一等公民工具检索，降低 Tool Calling 延迟并提供更精准的个性化上下文。
  • 反思流水线统一 LLM 客户端：记忆写入、反思、查询统一使用记忆模块封装的 LLM 客户端，模型获取更一致，故障与漂移下更稳定。
  • 长列表消息独立附属节点：需剪枝的长列表/清单型消息自动生成独立附属节点并关联用户与文档实体，原文检索准确率显著提升。
  • 剪枝去重与 Token 节省：剪枝结果落到记忆消息层持久化，相同内容不再重复剪枝，重复摄入场景下 LLM 调用与 Token 消耗明显下降。

2. 记忆生命周期与遗忘引擎 🕰️

  • 记忆节点存储上限与自动遗忘：新增可配置的存储上限，超阈值自动触发遗忘策略，控制总量、避免长期使用下检索性能衰减。
  • 前端直接删除记忆节点：记忆管理页面支持直接删除记忆节点，人工维护与审计式清理效率显著提升。
  • 遗忘引擎按活跃度决策：遗忘引擎现在能感知每条记忆的诞生时间与被使用次数，自动清理长期无人问津的记忆，保留经常被引用的核心内容。
  • 剪枝空值容错与默认开启迁移：剪枝允许 QA 素材为空，迁移脚本默认开启剪枝，MCP、API 与工作流自定义输入无需手动切换即可获得更聚焦、更省 Token 的记忆萃取。

3. 知识库 📚

  • QA 数据一键导出 CSV：文档列表页新增"导出 QA"按钮，支持问答数据全量导出为 CSV，便于批量备份、离线编辑与跨系统迁移。
  • PPT 解析升级至 MinerU V3：PPT/PPTX 解析管道对接 MinerU V3，同步补齐 PDF 降级流程回调，格式感知逻辑与容错机制更强。
  • 单库检索重排与过滤逻辑优化：混合检索去除双重重排并支持自定义重排模型，分词检索关闭全局阈值过滤，检索速度与召回完整度兼得。

4. 应用与工作流 ⚙️

  • License 管理平台：License 管理平台正式上线，统一管理权限、客户、授权的全生命周期。
  • 工作流/智能体迁移自动绑定记忆配置：跨空间迁移后，"记忆提取""记忆存储"节点自动绑定目标空间默认配置，无需手动重选。
  • 内置 PostgreSQL 查询工具：内置数据库查询工具原生支持 PostgreSQL 连接与在线查询，工作流、智能体可直接调用，无需二次开发。
  • 私有化 MinerU 支持：工作流与智能体支持配置并调用私有化部署的 MinerU 服务，满足本地化数据安全管控要求。
  • HTTP 节点"导入 cURL"：工作流 HTTP 请求节点新增"导入 cURL"功能，一步将 cURL 命令翻译为完整 HTTP 请求配置。
  • SSO 工作空间默认模型回退：租户通过 SSO 建工作空间接口初始化时若未传模型参数，自动复用全局默认模型配置，保障空间记忆能力正常初始化。

5. 性能与稳定性 ⚡

  • 对话全链路耗时优化：重构节点数据库连接逻辑，减少连接开销、缩短整体响应时长，高频聊天与记忆查询卡顿显著减少。

🧭 未来展望

走到 v0.3.12，MemoryBear 的记忆不再只是"存起来、查得到"——它被看见、被称重、被优雅地忘却。反思引擎开始理解自己的边界情况；遗忘引擎依据真实活跃度指标做出判定；记忆的价值权重通过前端可视化面板走到用户眼前。这些改动共同推动产品从"功能齐备"迈向"自感知、自愈合"的记忆底座——一个可以承担生产负载、无需持续手动干预的记忆系统。

与此同时，极速检索与用户画像作为一等公民接入 Agent 工具链，叠加内置 PostgreSQL 查询工具、私有化 MinerU 支持、工作流 cURL 导入等能力，标志着 MemoryBear 正在从"记忆服务"演进为"面向上下文感知型 Agent 的基础设施层"。License 管理平台补齐了企业级最后一环，把权限、客户、授权的全生命周期收拢到统一平台之内。产品正朝着记忆、知识、工作流编排三位一体、并具备生产级治理能力的方向持续演进。

下个版本我们将围绕几个方向继续打磨。开放接口将迎来更精细的用量与频次管理，让不同规模的团队都能安心接入；记忆的成长与遗忘会被串成一条可追溯的时间线，每一次记忆的诞生、演化与消退都清晰可见；短期记忆的检索时效将进一步提升，让当下的对话被更快理解、更准引用。此外，我们会引入业界公认的长期记忆评测标准，用更高的尺度衡量记忆能力，并持续优化知识库在高并发与纯图片场景下的表现。体验侧，工作流与对话流将支持一问多答的流式回复；参数配置在自运行、分享与 API 调用三种场景下也会更加一致，更贴合真实使用习惯。明镜既已擦亮，下一步是让它看得更清、答得更快、忘得更明。

Assets 2
Loading
0 Join discussion