Repository navigation
Releases
JiuwenMemory v0.2.1
🎉 JiuwenMemory v0.2.1 版本发布说明
发布日期:2026 年 11 月 13 日
openJiuwen 社区正式发布 JiuwenMemory v0.2.1。本版本以 v0.2.0 的架构内核为基线,重点补齐树状分层记忆、Schema 时序与实体双轴检索、文档记忆、安全加固与定时演进等能力,并新增 MCP 接入面、上下文感知查询改写与 agent 方向适配;同时集中修复了一批接入层、存储与生命周期相关的正确性问题。
🚀 新特性
树状分层记忆
- 跨源结构关系:在既有「披露 / 多模态粒度 / 认知分层」三轴之外引入第四轴——以
HierarchyRef表达跨MemoryUnit的父子包含,父节点作为可检索概要、子节点作为可按需展开的证据。 - 显式建树与区间重建:通过
evolve的HIERARCHY模式显式建树,支持复用完整旧树与按时间区间重建(replace_in_span以叶权威为边界)。 - 分层检索:
search支持按父侧 role 优先召回、祖先上卷(rollup)与向下展开(expand_depth),展开选子与max_tokens共用预算;支持默认 TimeSpan 查询预设与已召回后代补位。 - 后台增量构建:支持固定 home 的后台自动增量建树、安全封口与故障保护;
Snapshot自动接入并在同步 infer 中继承。首期结构 kind 为TIME,其余 kind 仅提供模型与协议。
Schema 时序与实体双轴检索
- TemporalEntity 只读视图:由已持久化的 Schema Property 按 canonical identity 组装实体时间线,不写入任何存储,保留 Source-first 兜底。
- 实体 / 属性两条独立召回路径:实体路径走实体 seed +
schema_entities召回后做实体内属性二次选择;属性路径绕过实体名额截断直达召回,再晚融合回填所属实体。 - 双时间轴:事件时间(
t_event)与系统有效时间([t_valid, t_invalid)/knowledge_as_of)独立计算,支持latest / snapshot / history / range模式。 - 默认关闭、请求级启用:受装配期硬门
globals.schema_temporal_enabled控制,单次查询通过RetrievalQuery.schema_temporal显式开启,未启用时不改变普通检索行为。
文档记忆
- Markdown 人类可读真源:开启
globals.write_document后,记忆按归属类别与 project 坐标落成USER.md/MEMORY.md/daily_memory/日期.md,可直接用编辑器阅读。 - 单文件影子索引:以单个 SQLite 文件承载 FTS5 倒排、sqlite-vec 向量与全量
unit_json,统一承接点查、检索与list;缺少 embedder / sqlite-vec 时优雅降级。 - 看门狗双向同步:监听用户手改 Markdown,以 unit 粒度增量同步回影子索引,并通过写窗口防护避免写入与对账之间的回环竞态;路径访问做
realpath + commonpath逃逸校验。 - md 文件级接口:新增
read_md_file/edit_md_file,按文件路径读写原文;edit_md_file采用继承式 SUPERSEDE,保留原 unit 元数据并维护版本链。
安全加固
- 认证内核:支持
dev/api_key/trusted三种认证方式与可信网关主体初始化入口,身份只经认证中间件注入,请求 payload 中的身份保留字段在契约边界直接拒绝。 - 静态加密:面向 KV / 文件系统等存储提供 ENC1 加密装饰器与统一装配开关,支持全存储后端静态加密与索引部署保护契约。
- 角色感知授权与代理:以 Grant / Delegation 为真源,按角色与空间两轴判定,支持 Agent 代用户操作。
- 审计完整性:可选的链式 HMAC(
chained_hmac)审计,提供内存 / SQLite 链式 Store 与受保护审计;未装配时显式返回unsupported。 - 本地媒体准入:视频导入按落点 Scope 与服务端
local_media_roots精确绑定读取范围,无绑定默认拒绝。
检索增强
- 上下文感知 Query Rewrite:显式开启后,结合调用方提供的近期上下文,把口语化 / 省略主体的 query 改写为自包含检索短语;仅增强软召回信号,不生成答案、不改变权限与硬过滤。
- 绝对日期两侧共享规则:查询侧解析时间窗、写入侧生成
t_event共用同一first_absolute_date规则,修复「查询某日期无法命中同日记忆」。
多形态接入
- MCP 接入面:MCP 工具与
MemoryAPI全量对齐(38/38),复用 HTTP/CLI 的契约校验与调用桥;支持 async 工具线程隔离、认证 fail-closed 与工具签名 / 契约锁测试。 - 接入层一致化:HTTP / CLI / MCP / SDK 统一补齐
as_of、mode、tier、channel等参数转发,统一错误语义,修复 legacy 入口参数静默丢弃与误报 500。 - Agent 方向适配:新增
agent_llmExtractor(中英文可配、抽取时同步判定memory_class)、AgentEngine 文档模式读路径,并更新 jiuwenswarm 适配层。
工具与工程
- 应用 ruff
UP规则迁移类型注解导入,新增 ruff lint 与单元测试 CI 工作流。
🛠 修复与稳定性
- 存储与生命周期:修复看门狗物理删除 SUPERSEDE 旧版行导致版本链历史丢失;修复
#开头正文被误判为标题引发的连坐删除;修复 lifecycle 治理未接文档真源;修复不可解码MemoryUnit记录中断读取;统一引擎时间比较语义;已配置后端依赖预加载提前至装配期。 - 检索与构建:修复 QueryParser
globals开关被 defaults 遮蔽;优化树状记忆召回过滤与子节点补位;大文本写路径预切分与 LLM prompt 护栏,修复同步add卡死 900s;统一 construction LLM 尝试次数语义;Fuser._normalize_weights对非法通道名抛ValidationError;本地 BGE 模型解析与不可用 reranker 降级。 - 接入层:修复远程 CLI 超时配置与错误处理;修复 MCP 服务生命周期上下文管理器;修复
_search/_delete/_add/_evolve参数校验与枚举解析;修复_update/_get参数遗漏;空内容消息过滤并保留 HTTPException 状态码。 - 安全与导入:修复 Markdown 标题日志敏感信息泄露;修复视频导入本地文件访问范围;核验加密卷初始化的 ext4 内部 journal;修复可选依赖导入吞噬内部 ImportError;按实例隔离 LLM 晚绑定端点键;默认以 USER 运行并支持显式 ROOT。
- 时序与调度:修复乱序消息写入的时序校验与反向失效处理;完善异步调度器关闭与后台线程回收,按需创建 daemon thread;修复 ingest 任务无法取消;修复上游限流被误报为不可重试的 500。
📚 兼容性说明
本版本在 v0.2.0 公开接口基础上保持能力与接口兼容,新增能力默认关闭或保持原有行为:树状记忆、Schema 时序检索与上下文改写均需显式开启,文档记忆与 dreaming 由配置开关控制;安全认证与加密按部署配置启用。升级前请结合实际部署形态核对配置项与安全默认值,并规划所需后端依赖(加密卷、审计 Store 等)。
以上就是 JiuwenMemory v0.2.1 的完整发布内容!感谢您使用 JiuwenMemory!我们期待您的反馈和建议!🚀
🎉 JiuwenMemory v0.2.1 Release Notes
Release Date: November 13, 2026
The openJiuwen community officially releases JiuwenMemory v0.2.1. Building on the v0.2.0 kernel as its baseline, this release focuses on tree-structured hierarchical memory, schema-temporal and entity dual-axis retrieval, document memory, security hardening, and scheduled evolution, while adding an MCP surface, context-aware query rewrite, and agent-oriented adaptation. It also fixes a batch of correctness issues across the access layer, storage, and lifecycle.
🚀 New Features
Tree-Structured Hierarchical Memory
- Cross-source structural relations: Adds a fourth axis on top of disclosure / multimodal granularity / cognitive tier —
HierarchyRefexpresses parent-child containment acrossMemoryUnits, using parents as retrievable summaries and children as on-demand evidence. - Explicit construction and interval rebuild: Build trees explicitly via
evolve'sHIERARCHYmode, with full-tree reuse and time-interval rebuild (replace_in_spanbounded by leaf authority). - Hierarchical retrieval:
searchsupports parent-role-prioritized recall, ancestor rollup, and downward expansion (expand_depth); expansion shares themax_tokensbudget, with default TimeSpan query presets and backfill of already-recalled descendants. - Background incremental construction: Fixed-home background incremental tree building with safe sealing and fault protection;
Snapshots attach automatically and are inherited in synchronous infer. The first structural kind isTIME; other kinds provide models and protocols only.
Schema-Temporal and Entity Dual-Axis Retrieval
- Read-only TemporalEntity view: Assembled from persisted Schema Properties by canonical identity; nothing is written to storage, preserving the Source-first fallback.
- Independent entity / property recall paths: The entity path recalls entity seeds plus
schema_entities, then selects properties within the entity; the property path bypasses entity quota truncation and late-fuses back to its entity. - Dual time axes: Event time (
t_event) and system valid time ([t_valid, t_invalid)/knowledge_as_of) are computed independently, supportinglatest / snapshot / history / rangemodes. - Off by default, enabled per request: Gated by
globals.schema_temporal_enabled; enabled per query viaRetrievalQuery.schema_temporal, with ordinary retrieval unchanged when disabled.
Document Memory
- Human-readable Markdown source of truth: With
globals.write_documentenabled, memories land asUSER.md/MEMORY.md/daily_memory/<date>.mdby class and project, directly readable in an editor. - Single-file shadow index: One SQLite file carries FTS5 full-text, sqlite-vec vectors, and full
unit_json, serving point reads, retrieval, andlist; degrades gracefully without an embedder / sqlite-vec. - Watchdog two-way sync: Hand edits to Markdown are synced back to the shadow index at unit granularity, with write-window protection against write/reconcile races and
realpath + commonpathpath-escape checks. - File-level md APIs: New
read_md_file/edit_md_fileoperate on raw content by file path;edit_md_fileuses inherited SUPERSEDE, preserving unit metadata and the version chain.
Security Hardening
- Authentication kernel:
dev/api_key/trustedmodes and a trusted-gateway principal initialization entry; identity is injected only by auth middleware, and reserved identity fields in request payloads are rejected at the contract boundary. - Encryption at rest: ENC1 encryption decorators and a unified assembly switch for KV / filesystem stores, with index deployment protection contracts.
- Role-aware authorization and delegation: Grant / Delegation as the source of truth, decided along role and space axes, supporting agent operations on behalf of a user.
- Audit integrity: Optional chained HMAC (
chained_hmac) audit with in-memory / SQLite chained stores and protected audit; explicitly reportsunsupportedwhen not assembled. - Local media admission: Video import binds read scope to the server's
local_media_rootsper target Scope; denied by default without a binding.
Retrieval Enhancements
- Context-aware query rewrite: When explicitly enabled, rewrites colloquial or subject-elided queries into self-contained retrieval phrases using caller-provided recent context; it only strengthens soft recall signals and never changes permissions or hard filters.
- Shared absolute-date rule: Query-side time-window parsing and write-side
t_eventgeneration share the samefirst_absolute_daterule, fixing date queries that missed same-day memories.
Multi-Form Access
- MCP surface: MCP tools fully align with
MemoryAPI(38/38), reusing the HTTP/CLI contract validation and invocation bridge; includes async tool thread isolation, fail-closed auth, and tool-signature / contract lock tests. - Access-layer consistency: HTTP / CLI / MCP / SDK uniformly forward
as_of,mode,tier,channel, and more; unified error semantics, fixing silent parameter drops and false 500s in the legacy entry. - Agent-oriented adaptation: New
agent_llmExtractor (bilingual, determinesmemory_classduring extraction), an AgentEngine document-mode read path, and updated jiuwenswarm adapter.
Tooling & Engineering
- Applied ruff
UPrules to annotation imports, and added ruff lint plus unit-test CI workflows.
🛠 Fixes & Stability
- Storage & lifecycle: Fixed watchdog physical deletion of superseded rows losing version-chain history; fixed
#-prefixed body lines misread as titles triggering cascading deletes; fixed lifecycle governance not wired to the document source of truth; isolated undecodableMemoryUnitrecords; unified engine time-comparison semantics; moved configured-backend preloading to assembly time. - Retrieval & construction: Fixed QueryParser
globalsswitches shadowed by defaults; improved tree recall filtering and child backfill; added large-text pre-chunking and prompt guards to fix synchronousaddhanging for 900s; unified ...
JiuwenMemory v0.2.0
🎉 JiuwenMemory v0.2.0 版本发布说明
发布日期:2026 年 09 月 21 日
openJiuwen 社区正式发布 JiuwenMemory v0.2.0。作为全新架构升级后的首个正式版本,JiuwenMemory 为 Agent 提供记忆抽取、构建、检索、演进与管理的全链路能力,既可部署为独立记忆服务,也可作为库嵌入任意 Agent 应用,全面支持云服务与独立 Agent 生态。
🚀 新特性
记忆抽取与写入
- 多类型记忆自动抽取:从对话与文本中自动抽取情景、语义、程序(procedural)、core 四类记忆,无需手工配置规则;支持自定义抽取模式与动态 prompt 策略,按场景定制抽取行为。
- 智能写入与批量写入:支持写入时可选过程记忆抽取和派生事实抽取;并提供
batch_add、batch_add_async批量写入记忆接口,支持一次提交多条记忆,保持单条语义不变,统一错误处理与结果对齐。 - 视频多模态记忆:支持视频内容的记忆构建与检索——视频流水线基于 DashScope ASR 完成音轨转写,结合多模态抽取能力从视频中提取记忆,并支持多模态记忆的混合检索召回。
AgentTeam 模式记忆
- 群体记忆构建与跨空间召回:支持 Agent Team(多智能体团队)场景下的集体记忆(collective memory)构建,团队成员的记忆可跨 Space 召回共享,为团队级协作提供统一记忆底座。
混合记忆检索
- 多路召回:支持关键词 / 向量混合召回,支持 L0/L1 分层索引下的多路并行召回。
- 多种融合策略:RRF(默认)/ 加权 RRF / 最大值融合(CombMAX)。
- 树形 metadata 过滤:AND/OR/NOT + 九种比较操作符,在 Milvus / Elasticsearch / pgvector 上完整下推,支持 as_of 历史时刻查询。
多租户与记忆治理
- Space 多租户隔离:基于 Space 的记忆隔离与全生命周期管理(创建 / 归档 / 导出 / 用量 / 策略 / 成员 / 清除)。
- 细粒度授权:
grant/revoke,授权持久化、软撤销、过期校验、跨组织拒绝。 - 治理能力:
inspect/trace/audit,审计日志可落盘、可按操作者与目标过滤。
多形态接入
- 两种接入形态统一映射到同一记忆内核:
- SDK:进程内嵌入,一行
assemble()即可获得纯内存离线栈; - HTTP Server:完整 REST verbs(add / batch_add / list / job / space / audit 等)。
- SDK:进程内嵌入,一行
可插拔存储与部署
- 六类存储抽象:KV / 全文 / 向量 / 图 / 融合 / 文件系统,统一 CRUD + search 契约,scope 五维原生隔离。
- 灵活部署:从纯内存离线栈到全后端生产部署平滑演进;Docker Compose 三套形态——local(纯本地)/ online(Redis + Milvus + ES)/ postgres(PostgreSQL + pgvector + ES),容器非 root 运行。
开放生态与可扩展性
- 可插拔组件:七类共享插件契约(Embedder / Chunker / Tokenizer / Normalizer / FeatureExtractor / LLM / Reranker);LLM Provider 适配 OpenAI 与 DashScope。
- 运行时配置:ConfigSource 晚绑定机制,能力开关、prompt、模型凭证、存储连接等配置支持运行时改值与实例热切换。
📚 兼容性说明
当前版本与 JiuwenMemory v0.1.x 版本在接口设计和具体实现上存在差异,但保持能力范围的兼容。v0.2.x 版本将作为后续长期演进版本,请使用 v0.1.x 版本的开发者合理规划版本迁移和适配工作。
以上就是 JiuwenMemory v0.2.0 的完整发布内容!感谢您使用 JiuwenMemory!我们期待您的反馈和建议!🚀
🎉 JiuwenMemory v0.2.0 Release Notes
Release Date: September 21, 2026
The openJiuwen community officially releases JiuwenMemory v0.2.0. As the first stable release after a full architectural upgrade, JiuwenMemory provides agents with end-to-end memory capabilities — extraction, construction, retrieval, evolution, and management. It can be deployed as a standalone memory service or embedded as a library into any agent application, fully supporting both cloud services and the independent agent ecosystem.
🚀 New Features
Memory Extraction & Writing
- Automatic multi-type memory extraction: Automatically extracts four memory types from conversations and text — episodic, semantic, procedural, and core — with no hand-crafted rules required. Supports custom extraction modes and dynamic prompt strategies to tailor extraction behavior per scenario.
- Smart & batch writing: Optional process-memory extraction and derived-fact extraction on write; provides
batch_addandbatch_add_asyncbatch-write APIs that submit multiple memories in one call while preserving single-write semantics, with unified error handling and aligned results. - Video multimodal memory: Supports memory construction and retrieval over video content — the video pipeline transcribes audio tracks via DashScope ASR, extracts memories from video using multimodal extraction, and supports hybrid retrieval and recall of multimodal memories.
AgentTeam Mode Memory
- Collective memory construction & cross-space recall: Supports collective memory construction for Agent Team (multi-agent team) scenarios. Team members' memories can be recalled and shared across Spaces, providing a unified memory foundation for team-level collaboration.
Hybrid Memory Retrieval
- Multi-path recall: Hybrid keyword / vector recall, with multi-path parallel recall over L0/L1 layered indexes.
- Multiple fusion strategies: RRF (default) / weighted RRF / CombMAX.
- Tree-structured metadata filtering: AND/OR/NOT plus nine comparison operators, fully pushed down on Milvus / Elasticsearch / pgvector, with support for as-of historical-time queries.
Multi-Tenancy & Memory Governance
- Space-based multi-tenant isolation: Memory isolation by Space with full lifecycle management (create / archive / export / usage / policy / members / purge).
- Fine-grained authorization:
grant/revokewith persistent grants, soft revocation, expiration checks, and cross-organization denial. - Governance capabilities:
inspect/trace/audit; audit logs can be persisted to disk and filtered by operator and target.
Multiple Access Forms
- Two access forms unified onto the same memory kernel:
- SDK: In-process embedding — a single
assemble()call yields a pure in-memory offline stack; - HTTP Server: Full REST verbs (add / batch_add / list / job / space / audit, etc.).
- SDK: In-process embedding — a single
Pluggable Storage & Deployment
- Six storage abstractions: KV / full-text / vector / graph / fusion / filesystem, with a unified CRUD + search contract and native five-dimensional scope isolation.
- Flexible deployment: Smooth evolution from a pure in-memory offline stack to full-backend production deployment. Three Docker Compose profiles — local (pure local) / online (Redis + Milvus + ES) / postgres (PostgreSQL + pgvector + ES) — with containers running as non-root.
Open Ecosystem & Extensibility
- Pluggable components: Seven shared plugin contracts (Embedder / Chunker / Tokenizer / Normalizer / FeatureExtractor / LLM / Reranker); LLM providers adapted for OpenAI and DashScope.
- Runtime configuration: ConfigSource late-binding mechanism — capability switches, prompts, model credentials, storage connections, and more support runtime updates and hot instance swapping.
📚 Compatibility Notes
This release differs from JiuwenMemory v0.1.x in API design and implementation details, but remains compatible in terms of capability coverage. The v0.2.x line will serve as the long-term evolution branch; developers on v0.1.x are encouraged to plan their migration accordingly.
That's everything in JiuwenMemory v0.2.0! Thank you for using JiuwenMemory — we look forward to your feedback and suggestions! 🚀
JiuwenMemory v0.2.0.beta1
🌟JiuwemMemory v0.2.0.beta1 版本发布说明
发布日期:2026 年 09 月 04 日
openJiuwen 社区发布 JiuwenMemory v0.2.0.beta1,作为全新架构升级后的首个版本预发布版,为 Agent 提供记忆抽取、构建、检索、演进与管理的全链路能力,支持独立为记忆服务,也可作为库嵌入任意 Agent 应用,全面支持云服务与独立Agent生态。
🚀 新特性
记忆抽取与写入
- 多类型记忆自动抽取:从对话与文本中自动抽取情景、语义、程序(procedural)、core 四类记忆,无需手工配置规则;支持自定义抽取模式与动态 prompt 策略,按场景定制抽取行为。
- 智能写入与批量写入:支持写入时可选过程记忆抽取和派生事实抽取;并提供
batch_add、batch_add_async批量写入记忆接口,支持一次提交多条记忆,保持单条语义不变,统一错误处理与结果对齐。
混合记忆检索
- 多路召回:支持关键词 / 向量 混合召回,支持 L0/L1 分层索引下的多路并行召回。
- 多种融合策略:RRF(默认)/ 加权 RRF / 最大值融合(CombMAX)。
- 树形 metadata 过滤:AND/OR/NOT + 九种比较操作符,在 Milvus / Elasticsearch / pgvector 上完整下推,支持 as_of 历史时刻查询。
多租户与记忆治理
- Space 多租户隔离:基于 Space 的记忆隔离与全生命周期管理(创建 / 归档 / 导出 / 用量 / 策略 / 成员 / 清除)。
- 细粒度授权:
grant/revoke,授权持久化、软撤销、过期校验、跨组织拒绝。 - 治理能力:
inspect/trace/audit,审计日志可落盘、可按操作者与目标过滤。
多形态接入
- 两种接入形态统一映射到同一记忆内核:
- SDK:进程内嵌入,一行
assemble()即可获得纯内存离线栈; - HTTP Server:完整 REST verbs(add / batch_add / list / job / space / audit 等);
- SDK:进程内嵌入,一行
可插拔存储与部署
- 六类存储抽象:KV / 全文 / 向量 / 图 / 融合 / 文件系统,统一 CRUD + search 契约,scope 五维原生隔离。
- 灵活部署:从纯内存离线栈到全后端生产部署平滑演进;Docker Compose 三套形态——local(纯本地)/ online(Redis + Milvus + ES)/ postgres(PostgreSQL + pgvector + ES),容器非 root 运行。
开放生态与可扩展性
- 可插拔组件:七类共享插件契约(Embedder / Chunker / Tokenizer / Normalizer / FeatureExtractor / LLM / Reranker);LLM Provider 适配 OpenAI 与 DashScope。
- 运行时配置:ConfigSource 晚绑定机制,能力开关、prompt、模型凭证、存储连接等配置支持运行时改值与实例热切换。
📚 兼容性说明
当前版本与 JiuwemMemory v0.1.x版本在接口设计和具体实现上存在差异,但是保持能力范围的兼容。v0.2.x版本将作为后续长期演进版本,请使用v0.1.x版本的开发者合理规划版本迁移和适配工作。
以上就是 JiuwemMemory v0.2.0.beta1 的完整发布内容!感谢您使用 JiuwemMemory!我们期待您的反馈和建议!🚀
JiuwenMemory v0.1.2
JiuwenMemory v0.1.2 Release Notes
发布信息
版本:v0.1.2
发布日期:2026-09-02
项目仓库:https://github.com/openJiuwen-ai/agent-memory
版本简述
本次 v0.1.2 版本新增记忆服务化 RESTful API、冲突处理与遗忘机制、Docker 一键部署、数据库 SSL 连接、国密算法存储加密与过期记忆批量治理六大企业级能力;同时修复 36 项问题,覆盖记忆 CRUD 接口参数校验与并发安全、向量存储检索一致性(ES/Milvus)、mem_meta 过期记忆任务系统(建表竞态、MySQL 启动期死锁、时区口径漂移等)、file 后端稳定性、GaussDB 方言适配与连接池配置,并完成服务化平台配置治理与两轮代码扫描整改;同步完善批量删除与记忆遗忘相关接口文档,全面提升记忆引擎在多存储后端下的数据一致性与生产环境运行稳定性。本里程碑关联的 44 项 issue 已全部关闭。
🚀 新增特性
- 记忆服务化能力
支持查询、监控、可视化等记忆服务化能力,对外提供 RESTful API,统一上层业务调用规范,降低记忆能力的接入与运维成本。(#259) - 支持一键部署,服务 Docker 化
提供 Docker 化一键部署方案,简化环境准备与服务安装流程,缩短从获取代码到服务可用的路径。(#258) - 数据库连接支持 SSL 证书
数据库连接新增 SSL 证书支持,保障数据传输链路安全,满足企业级安全合规要求。(#235) - 记忆存储加密支持国密算法
记忆存储加密算法扩展支持国密算法(SM 系列),满足国内合规场景下的数据加密要求。(#255) - 冲突处理与遗忘机制支持
支持双时序(创建/更新时间)、多种遗忘策略可配置,记忆版本可追溯与可视化,为记忆治理提供完善的冲突消解与生命周期管理能力。(#249) - 过期记忆批量治理与阈值告警
新增基于用户维度记忆条数的多级阈值告警、过期记忆批量删除与 VACUUM 空间回收机制,实现记忆存量的可持续治理,避免记忆无限膨胀。(#104)
📄 文档改进
- 补充批量删除过期记忆接口文档
将"批量删除过期记忆"涉及的 4 个接口补充至官方 docs,完善接口说明与参数定义。(#166) - 同步记忆遗忘特性接口文档
将遗忘记忆特性新增及修改的接口同步至 docs,消除文档与代码不同步的问题。(#282)
🛠️ 问题修复
记忆 CRUD 接口校验与并发安全修复
- 修复
update_mem_by_id对不存在的mem_id执行"先删后写"返回 200 但数据未持久化的问题
补充mem_id存在性校验,杜绝假成功响应,保证更新操作的原子性与可观测性。(#251) - 修复
update/delete_mem_by_id对不存在或空mem_id无校验、盲走中期记忆 upsert/delete 的问题
增加入参前置校验,消除假成功污染中期向量库或以误导性[158004]报错的情况。(#248) - 修复 simple 后端
update_memories为非原子 delete-then-add,并发更新同一mem_id必现[158004] not found的问题
改造更新流程保证原子性,并发更新场景不再报记忆不存在。(#269) - 修复
get_user_mem_by_page响应缺失source_id字段导致记忆溯源链路断裂的问题
补全响应字段,恢复完整的记忆溯源能力。(#231) - 修复
add_messages对空content返回 500 的问题
增加参数合法性校验,空content按预期返回明确的业务错误而非服务端异常。(#213) - 修复
get_message_by_id对不存在 ID 返回 500 的问题
按契约返回 200 且"found": false,与查询类接口行为保持一致。(#211) - 修复
get_user_mem_by_page的page_size入参缺失校验的问题
负值不再透传至底层引擎,避免 400 错误被误包装为 500。(#243) - 修复 HTTP
/add_messages/接口缺少session_id参数的问题
补齐会话标识参数,保证消息与会话正确关联。(#242) - 修复写入后分页列表数据不可见、删除全部记忆后仍可检索、更新后检索仍为旧值的问题
修正记忆增删改后的索引刷新与可见性逻辑,保证读写一致性。(#240) - 修复服务化平台使用过程中发现的多项问题
集中修复服务化平台试用反馈的功能缺陷,提升服务化链路整体可用性。(#65)
向量存储与检索一致性修复
- 修复
batch_delete_mem不删除 ES 向量库中期记忆的问题
补齐 ES 中期记忆的删除链路,避免批量删除后向量库残留脏数据。(#250) - 修复
get_user_mem_by_page全类型分支 filters 对中期记忆子集失效的问题
修正过滤条件下推逻辑,EQ(blacklisted=true)等查询不再被中期记忆数据污染。(#246) - 修复 Milvus 搜索时
output_fields未正确包含content字段的问题
补全查询输出字段,恢复所有middle_term_memory搜索的正确返回。(#212) - 修复 Elasticsearch 后端
middle_term_memory写入失败的问题
_build_mappings不再硬编码dynamic: "strict",未声明字段不再被拒绝写入。(#234) - 修复
get_user_mem_by_page分页跨页返回重复记录并跳过部分记录的问题
为simple_memory_index排序增加稳定次级键,不再依赖底层向量库不稳定的返回顺序。(#39) - 修复
MilvusVectorStore.list_collection_names绕过懒加载导致空库查询抛AttributeError的问题
统一走get_client()懒加载入口,空库场景可正常返回空列表。(#76) - 修复
delete_mem_by_user_id/delete_mem_by_scope不清理retrieve_historyKV 的问题
对齐删除契约,与delete_mem_by_id行为保持一致,避免检索历史残留。(#26) - 修复 file 后端 search 混合检索中 FTS 停用词污染召回排序的问题
修正全文检索分词处理,恢复混合检索的召回质量。(#36) - 修复
async_gaussdb跨事件循环连接池绑定及 ES 向量存储 schema 不匹配的问题
修正异步连接池的生命周期管理,对齐 ES 向量存储 schema 定义。(#239)
记忆遗忘 / file 后端稳定性修复
- 修复 file 后端调用
search_user_mem或get_user_mem_by_page直接抛异常的问题
补齐 file 存储后端的检索与分页实现,接口在 file 后端下可正常使用。(#72) - 修复
file_memory_indexsqlite 写锁竞争导致/health探针超时的问题
优化写事务持锁范围,长稳测试中 health 探针成功率恢复正常水平。(#37)
mem_meta 过期记忆任务系统修复
- 修复独立库首请求撞建表竞态的问题
_init_db不再 fire-and-forget,保证建表完成后才处理请求,消除首请求 refresh 500。(#210) - 修复 mysql 组合启动期死锁的问题
消除cleanup_zombie_tasks后台 UPDATE 与首请求 refresh INSERT 在mem_meta_task表上的死锁(错误码 1213)。(#206) - 修复 mem_meta 表定义使用无长度 String、mysql 建表即
CompileError的问题
修正表定义并增加启动期建表失败感知,避免四接口全 500 且无告警。(#199) - 修复并发提交 refresh 防重竞态的问题
将 check-then-insert 改造为原子操作,双发请求不再同时 202 accepted 产生双任务。(#193) - 修复僵尸任务
started_at=NULL不清理导致batch_delete永久阻塞的问题
补齐僵尸任务识别与回收逻辑,批量删除任务不再被异常任务卡死。(#191) - 修复
force=true跳过运行中任务检查导致并发冲突的问题
强制提交时同样校验运行中任务,避免同 scope 并发任务互相踩踏。(#190) - 修复时间口径的三个问题
接口时间显式声明时区(替代 UTC 裸串)、替换弃用的utcnow()写法、修正记忆年龄计算时区混用导致的 30 天阈值 +8h 漂移。(#189) - 修复
batch_deletescope 级删除失败被任务级状态掩盖的问题
失败信息不再被终态 completed 吞掉,failed_count如实反映各 scope 的删除结果。(#180)
数据库与 GaussDB 适配修复
- 修复 GaussDB 连接池缺少
pool_recycle/pool_size配置的问题
补全连接池参数,十万级数据量下expired_memorys接口不再因连接失效返回 500。(#192) - 修复数据库使用 GaussDB 时 SQL 不支持现有
on conflict语句的问题
按 GaussDB 方言改写冲突处理 SQL,保障 GaussDB 部署形态下功能可用。(#238) - 修复
gauss_dialect导入路径写错的问题
修正模块导入路径,消除 GaussDB 方言加载失败。(#237)
服务化平台配置与代码质量修复
- 修复服务化模板配置硬编码大模型的问题
删除硬编码的模型配置,模型选择交由用户配置项管理,避免误导与污染。(#153) - 修复
ModelClientConfigdescription 中出现公司名称敏感词的问题
清理配置描述中的敏感词信息。(#148) - 完成 Python 内核代码逻辑扫描整改
依据静态扫描结果修复代码逻辑隐患,提升内核代码质量。(#194) - 完成服务化平台代码扫描整改
依据扫描结果修复服务化平台代码问题,提升服务化代码质量。(#188)
以上就是 JiuwenMemory v0.1.2 的完整更新内容!感谢您使用 JiuwenMemory!我们期待您的反馈和建议!🚀
JiuwenMemory v0.1.2 Release Notes (English)
Release Info
Version: v0.1.2
Release Date: Sep 2, 2026
Repository: https://github.com/openJiuwen-ai/agent-memory
Version Overview
This v0.1.2 release delivers six enterprise-grade capabilities: memory service with RESTful API, conflict handling & forgetting mechanisms, Docker-based one-click deployment, SSL-certified database connections, SM-series (Chinese national cryptography) storage encryption, and batch governance of expired memories with multi-level threshold alerts. It also resolves 36 defects covering memory CRUD parameter validation and concurrency safety, vector storage and retrieval consistency (ES/Milvus), the mem_meta expired-memory task system (table-creation race conditions, MySQL startup deadlocks, timezone drift, etc.), file backend stability, GaussDB dialect adaptation and connection pool configuration, along with service-platform configuration governance and two rounds of code-scan remediation. Documentation for the batch-delete and memory-forgetting APIs has also been completed. All 44 issues in this milestone are now closed.
🚀 New Features
- Memory-as-a-Service
Support query, monitoring and visualization capabilities for memory services via RESTful API, standardizing upstream integration and reducing the cost of adopting and operating memory capabilities. (#259) - One-Click Deployment with Docker
Provide a Dockerized one-click deployment solution, simplifying environment preparation and service installation from code to a running service. (#258) - SSL Certificate Support for Database Connections
Add SSL certificate support for database connections to secure data-in-transit and meet enterprise security compliance requirements. (#235) - SM-Series National Cryptography Support for Memory Encryption
Extend memory storage encryption to support Chinese national cryptography algorithms (SM series) for domestic compliance scenarios. (#255) - Conflict Handling & Forgetting Mechanisms
Support dual timestamps (created/updated), configurable multi-strategy forgetting policies, and traceable, visualizable memory versions, providing complete conflict resolution and lifecycle management for memory governance. (#249) - Batch Governance of Expired Memories with Threshold Alerts
Introduce multi-level threshold alerts based on per-user memory counts, batch deletion of expired memories, and VACUUM space reclamation, enabling sustainable memory governance and preventing unbounded memory growth. (#104)
📄 Documentation Improvements
- Batch Delete APIs Documentation
Document the 4 APIs involved in "batch delete of expired memories" in the official docs, completing API descriptions and parameter definitions. (#166) - Memory Forgetting APIs Documentation Sync
Sync the newly added and modified APIs of the memory forgetting feature into the docs, eliminating the doc-code inconsistency. (#282)
🛠️ Bug Fixes
Memory CRUD Validation & Concurrency Safety Fixes
- Fix
update_mem_by_idreturning 200 without persisting data for non-existentmem_idduring "delete-then-write"
Addmem_idexistence validation to eliminate false-success responses and ensure atomic, observable updates. (#251) - Fix missing validation on empty or non-existent
mem_idinupdate/delete_mem_by_idblindly hitting middle-term memory upsert/delete
Add pre-parameter validation to prevent false successes polluting the middle-term vector store or misleading[158004]errors. (#248) - Fix non-atomic delete-then-add in simple backend
update_memoriescausing inevitable[158004] not foundon concurrent updates to the samemem_id
Rework the update flow for atomicity; concurrent updates no longer report missing memories. (#269) - Fix missing
source_idfield inget_user_mem_by_pageresponse breaking the memory traceability chain
Restore the complete memory traceability capability. (#231) - Fix
add_messagesreturning 500 on emptycontent
Validate parameter legitimacy; emptycontentnow returns an explicit business error instead of a server exception. (#213) - Fix
get_message_by_idreturning 500 for non-existent IDs
Return 200 with"found": falseper the API contract, consistent with other query endpoints. (#211) - Fix missing validation for the
page_sizeparameter ofget_user_mem_by_page
Negative values are no longer passed through to the underlying engine, avoiding 400 errors being mis-wrapped as 500. (#243) - Fix missing
session_idparameter in the HTTP/add_messages/API
Add the session identifier to correctly associate messages with sessions. (#242) - Fix data invisible in paginated lists after writing, still retrievable after deleting all memories, and stale values returned after updates
Correct index refresh and visibility logic after memory create/update/delete to guarantee read-write consistency. (#240) - Fix multiple issues found during service platform usage
Consolidated fixes for functional defects reported during service ...
JiuwenMemory v0.1.1
JiuwenMemory v0.1.1 Release Notes
发布信息
版本:v0.1.1
发布日期:2026-07-18
项目仓库:https://github.com/openJiuwen-ai/agent-memory
版本简述
本次 v0.1.1 版本新增记忆 MCP 标准化服务、多代码智能体适配、文件系统持久化存储三大核心能力;同时修复 Milvus 检索无限阻塞、并发写入数据丢失、记忆角色字段丢失、变量更新接口静默失败等 5 项线上稳定性问题,统一向量存储超时规范、完善异常日志与边界测试用例,大幅提升记忆引擎存储扩展性与运行稳定性。
🚀 新增特性
- 新增记忆 MCP 标准化服务
对外提供统一 MCP 记忆服务能力,标准化记忆读写、检索、持久化接口,统一上层智能体调用规范,降低多业务、多 Agent 接入成本,配套完整参数校验、标准出入参定义与全量测试用例覆盖。(#80) - 支持多款代码智能体原生接入记忆引擎
完成 Codex、ClaudeCode、OpenCode 等代码类 Agent 适配改造,代码编写、调试场景可自动复用会话记忆、长期记忆,打通代码智能体上下文记忆链路,插件层完成全量兼容性适配验证。(#89) - 新增文件系统存储持久化方案
拓展底层存储介质,新增本地文件系统存储实现,支持无向量数据库环境下独立运行记忆模块;支持记忆分片落盘、定时持久化、本地文件读取恢复,轻量化部署场景无需依赖 Milvus 等向量引擎,降低单机轻量化部署门槛。(#82)
🛠️ 问题修复
向量存储阻塞类问题修复
- 修复 Milvus 向量查询无超时配置导致进程无限阻塞问题
为 Milvus 全量 RPC 操作(检索、插入、更新、删除)增加超时参数,区分连接超时与单次请求超时;服务无响应时主动抛出异常,避免测试套件、线上服务事件循环卡死。(#50) - 修复图结构混合检索
hybrid_search未配置超时,引发服务永久挂起
统一图记忆检索链路超时规范,补齐超时截止时间配置,消除向量检索无时限导致死锁风险。(#33)
记忆数据一致性与接口异常修复
- 修复同用户 / 会话高并发写入时消息数据静默丢失
重构并发写入锁机制,增加写入前置校验与事务兜底逻辑,通过并发压测验证无数据覆盖丢失,保障高并发记忆数据完整性。(#39) - 修复中长期记忆转换丢失 role 字段,助手消息错误存入长期记忆
补全消息角色存储逻辑,完整区分 user/assistant/system 消息类型,标准化分层记忆角色识别规则,避免记忆抽取逻辑错乱。(#32) - 修复 LongTermMemory 变量更新接口变量不存在时无提示静默失败
增加入参前置校验,变量不存在时主动抛出清晰业务异常并打印错误堆栈日志,替代原有无日志静默失败逻辑,提升问题排查效率。(#44)
JiuwenMemory v0.1.1 Release Notes (English)
Release Info
Version: v0.1.1
Release Date: 2026-07-18
Repo: https://github.com/openJiuwen-ai/agent-memory
Version Overview
This v0.1.1 release delivers three core capabilities: standardized Memory MCP Service, native integration for multiple code agents and filesystem-based persistent storage. It also resolves 5 critical stability defects including infinite blocking during Milvus retrieval, silent data loss under concurrent writes, missing role fields in memory conversion and silent failures of variable update APIs. We standardized timeout rules for vector storage, improved exception logging and supplemented boundary test cases to greatly enhance storage scalability and runtime stability of the memory engine.
🚀 New Features
- Add standardized Memory MCP Service
Provide unified MCP service for memory read, write, search and persistence. Standardize API specs for all upstream agents to lower integration cost for multi-business and multi-agent scenarios, with complete request validation, standard schema and full test coverage.(#80) - Native memory integration for multiple code agents
Fully adapt Codex, ClaudeCode, OpenCode and other code agents. Code writing & debugging workflows can automatically reuse session memory, long-term memory, unifying context memory pipeline for code agents with full plugin compatibility verification.(#89) - Add filesystem persistent storage support
Expand underlying storage medium with native local filesystem implementation. The memory module can run independently without vector databases. It supports sharded memory dumping, scheduled persistence and local file recovery, removing dependency on Milvus and lowering barriers for lightweight single-machine deployment.(#82)
🛠️ Bug Fixes
Vector Store Blocking Fixes
- Fix infinite process blocking caused by missing timeout on Milvus RPC calls
Add explicit timeout config for all Milvus RPC operations including search, insert, upsert and delete. Separate connection timeout and single-request timeout, throw explicit exceptions when Milvus unreachable to prevent test suite & online service loop hang.(#50) - Fix permanent service hang of graph hybrid_search without timeout config
Align unified timeout rules for all graph memory retrieval paths, add deadline limits to eliminate deadlock risk.(#33)
Memory Consistency & API Error Fixes
- Fix silent message loss under high concurrent writes for same user/scope
Refactor concurrent write lock mechanism with pre-write validation and transaction fallback. Verified by concurrent stress test to avoid message overwrite loss and guarantee data integrity.(#39) - Fix missing role field during medium-long memory conversion, leading to wrong assistant message storage
Complete role persistence logic to distinguish user/assistant/system messages, standardize role recognition for layered memory and prevent incorrect memory extraction.(#32) - Fix silent failure of LongTermMemory update_variables API when target variable not exists
Add pre-parameter validation, throw clear business exceptions with error stacks instead of silent failures to accelerate troubleshooting.(#44)
JiuwenMemory v0.1.0
🎉 JiuwenMemory v0.1.0 Release Note
发布日期:2026 年 7月 1 日
JiuwenMemory 正式开源!作为一款将 AI 记忆从被动信息存储重构为可治理、可跨平台共享、可自我演化核心数据资产的记忆引擎,它致力于让 Agent 真正具备"记住用户、理解用户、服务用户"的能力。
"模型能力决定 Agent 的'智力下限',记忆系统决定 Agent 的'体验上限'。"
本次 v0.1.0 发布作为项目的首个开源版本(Beta 状态),带来了分层记忆体系、Auto Dreaming 睡时巩固、记忆涡轮加速、知识图谱记忆、双维度解耦适配层以及丰富的存储后端体系等核心能力,构建完整的记忆提取、存储、检索和管理闭环。
🌟 JiuwenMemory v0.1.0 版本更新详情
发布日期:2026 年 6 月 30 日
发行说明:JiuwenMemory v0.1.0
🚀 全新特性
🧠 记忆构建和演化
- 分层记忆体系 (L0–L3):四层渐进架构(L0 原始信息 → L1 摘要记忆 → L3 结构化记忆 → L3 用户画像),各层独立持久存储,信息密度逐级放大,彻底解决长对话记忆丢失与偏好覆盖问题。支持用户画像、语义记忆、情景记忆、变量、摘要多种类型记忆自动提取,灵活配置自定义变量与禁止变量,精准匹配用户需求。
- Auto Dreaming(睡时记忆巩固):借鉴认知神经科学三阶段睡眠范式(浅睡筛选 → REM 提取归类 → 深睡去重消解),后台守护定时触发、忙碌退避与断点续扫,Token 开销线性可控。
- MemoryTurbo加速:对话瞬间写入缓存即完成更新,后台异步调度记忆提取;小模型按话题合并对话后一组提取,保证连贯性并大幅摊销大模型调用次数。
- Graph Memory(知识图谱记忆):支持对话/文档/JSON 多来源写入,LLM 自动抽取实体与关系,支持合并去重、图结构检索与 BFS 扩展,实体/关系/Episode 并行检索加 rerank 精准定位。
🔍 记忆检索和存储
- 语义检索与冲突检测:跨记忆类型统一向量语义检索,
MemUpdateChecker通过 LLM 分析语义冲突智能决定 ADD/DELETE 策略,LLM 输出 UPDATE/DELETE 指令经语义校验后执行,确保记忆一致与操作可控。 - 存储后端体系:全栈覆盖 KV(InMemoryKV/ShelveStore/DbBasedKV/Redis)、向量(ChromaDB/Milvus/Elasticsearch/GaussVector)、关系型(SQLite/PostgreSQL/MySQL/GaussDB)、消息(SqlMessageStore)、图(Milvus GraphStore)五大存储类别,适配本地单机到云端集群全场景。
- 数据迁移框架:支持 KV/向量/SQL/消息/索引的版本化 schema 迁移与跨 BaseMemoryIndex 实例批量数据迁移,操作注册表支持自定义迁移扩展。
🔌 生态接入与扩展
- REST API 服务与 OpenClaw 插件:FastAPI 完整 REST API + bearer-token 认证快速接入任意后端;OpenClaw JavaScript 生命周期插件自动记忆存储与召回,零配置即用。
🔒 安全
- 安全与并发控制:AES-256-GCM 透明加密记忆数据与 API Key 保障隐私;基于 KV 存储的分布式锁确保多实例并发一致;按
scope_id独立配置 LLM/嵌入模型/提取规则并加密存储,实现多租户安全隔离。
📊 性能表现
在 LoCoMo 评测集上,JiuwenMemory 以插件形式接入 OpenClaw:
- 准确率提升 15%
- 记忆问答与添加环节 Token 消耗降低 60%+
🗺️ 演进原则
JiuwenMemory 路线图遵循以下原则持续演进:
- 先解耦后扩展 — 在适配层完成 Agent 平台与记忆引擎的彻底解耦后,沿 Plugin 和 Provider 两个维度独立扩展
- 先治理后生长 — 冲突检测、遗忘机制、版本溯源等治理能力优先于多模态等生长能力
- 先个体后群体 — 单 Agent 记忆成熟后再演进至多 Agent 群体记忆共享
- 先本地后边云 — 本地存储体系稳定后逐步扩展至边云协同架构
最后感谢所有参与内测和提供反馈的开发者!JiuwenMemory 的开源标志着我们向"让 Agent 真正具备记忆能力"迈出了坚实的一步。欢迎大家在 GitCode 上提交 Issue 或 PR。
相关资源
- 官网:https://www.openjiuwen.com
- 代码仓库:https://gitcode.com/openJiuwen/jiuwenmemory
- 文档中心:https://www.openjiuwen.com/docs-page
🎉 JiuwenMemory v0.1.0 Official Open Source Release!
JiuwenMemory is now officially open source! As a memory engine that restructures AI memory from passive information storage into governable, cross-platform shared, and self-evolving core data assets, it empowers Agents with the ability to truly "remember, understand, and serve users."
"Model capabilities determine the 'lower bound of intelligence' for Agents; memory systems determine the 'upper bound of experience.'"
This v0.1.0 release marks the first open-source version (Beta status), introducing core capabilities such as a layered memory system, Auto Dreaming sleep-time consolidation, the MemorTurbo, Graph Memory, a dual-dimension decoupled adapter layer, and a comprehensive storage backend system — establishing a complete closed loop for memory extraction, storage, retrieval, and management.
🌟 JiuwenMemory v0.1.0 Release Details
Release Date: June 30, 2026
Release Notes: JiuwenMemory v0.1.0
🚀 Key Features
🧠 Memory Construction and Evolution
- Layered Memory System (L0–L3): Four-tier progressive architecture (L0 Raw Information → L1 Summary Memory → L2 Structured Memory → L3 User Profile) with independent persistent storage per layer, progressively increasing information density and resolving long-dialogue memory loss and preference overwrite issues. Supports automatic extraction of multi memory types — UserProfile, SemanticMemory, EpisodicMemory, Variable, and Summary — with flexible custom variable definitions and prohibited variable configurations for precise user need matching.
- Auto Dreaming (Sleep-Time Memory Consolidation): Three-stage sleep paradigm inspired by cognitive neuroscience (Light Sleep screening → REM extraction & classification → Deep Sleep deduplication & conflict resolution), with background daemon scheduling, busy-retreat deferment, and checkpoint-based incremental scanning for linearly controllable Token costs.
- MemoryTurbo: Raw conversations write to cache instantly for immediate updates while memory extraction runs asynchronously in the background; small model merges conversations by topic for group extraction, ensuring coherence and significantly amortizing large model call costs.
- Graph Memory (Knowledge Graph Memory): Supports CONVERSATION/DOCUMENT/JSON multi-source writing, LLM-powered automatic entity and relationship extraction with merge deduplication, graph-structure retrieval, BFS expansion, and parallel Entity/Relationship/Episode retrieval with rerank scoring for precise knowledge localization.
🔍 Memory Retrieval and Storage
- Semantic Retrieval & Conflict Detection: Unified cross-type vector semantic retrieval;
MemUpdateCheckeruses LLM to analyze semantic conflicts and intelligently decide ADD/DELETE strategies; LLM outputs UPDATE/DELETE directives validated via semantic checks before execution, ensuring memory consistency and controllable operations. - Storage Backend System: Full-stack coverage across five storage categories — KV (InMemoryKV/ShelveStore/DbBasedKV/Redis), Vector (ChromaDB/Milvus/Elasticsearch/GaussVector), Relational (SQLite/PostgreSQL/MySQL/GaussDB), Message (SqlMessageStore), and Graph (Milvus GraphStore) — adapting from local single-node to cloud cluster scenarios.
- Data Migration Framework: Supports versioned schema migration for KV/vector/SQL/message/index stores and cross-BaseMemoryIndex batch data migration, with an operation registry for custom migration extensions.
🔌 Ecosystem Integration and Extension
- Adapter Layer — Dual-Dimension Decoupling: Plugin dimension (hook-based memory injection, supporting JiuwenSwarm/OpenClaw/openJiuwen) and Provider dimension (unified
MemoryProviderinterface with JiuwenMemory/Mem0/openViking/AgentArts) extend independently, enabling N × M free combination. - REST API Service & OpenClaw Plugin: FastAPI complete REST API + bearer-token authentication for rapid backend integration; OpenClaw JavaScript lifecycle plugin for automatic memory storage and recall — zero-config, ready to use.
- Multi-Model LLM Clients: Supports OpenAI, DashScope, DeepSeek, SiliconFlow, OpenRouter, InferenceAffinity, IntelliRouter, and more — flexible selection of the optimal inference engine.
🔒 Security
- Security & Concurrency Control: AES-256-GCM transparent encryption for memory data and API Keys to safeguard privacy; distributed locking based on KV storage for multi-instance concurrency consistency; independent LLM/embedding/extraction rule configuration per
scope_idwith encrypted storage for multi-tenant security isolation.
📊 Performance
On the LoCoMo benchmark, JiuwenMemory was integrated with OpenClaw as a plugin:
- Accuracy improved by 15%
- Token consumption for memory Q&A and addition reduced by 60%+
🗺️ Evolution Principles
JiuwenMemory's roadmap follows these principles for continuous evolution:
- Decouple First, Then Extend — Complete decoupling of Agent platforms and memory engines in the adapter layer, then extend independently along Plugin and Provider dimensions
- Govern First, Then Grow — Governance capabilities (conflict detection, forgetting mechanisms, version traceability) precede growth capabilities (multimodal, etc.)
- Individual First, Then Group — Mature single-Agent memory before evolving to multi-Agent group memory sharing
- Local First, Then Edge-Cloud — Stabilize local storage systems before progressively extending to edge-cloud collaborative architecture
Special thanks to all developers who participated in internal testing and provided feedback! The open-sourcing of JiuwenMemory marks a solid step toward our vision of "empowering Agents with true memory capabilities." We welcome everyone to submit Issues or PRs on GitCode.
Related Resources
- Official Website: https://www.openjiuwen.com
- Repository: https://gitcode.com/openJiuwen/jiuwenmemory
- Documentation: https://www.openjiuwen.com/docs-page