← Back to Skills Marketplace
774
Downloads
0
Stars
4
Active Installs
89
Versions
Install in OpenClaw
/install unified-memory
Description
统一记忆系统 - AI Agent 专用记忆系统，支持 Context Tree、智能摘要、知识图谱、工作流引擎。零依赖，完整对标 QMD/MetaGPT
Usage Guidance
This package is internally inconsistent and should be treated cautiously. Steps to consider before installing:
- Do not install into a production agent or your primary account without verifying provenance. Ask the publisher for a canonical source repo and signatures.
- Inspect package.json, scripts, and the large merged-src.js locally (already included) to confirm behavior, especially openclaw hooks and scripts like deploy-atomic-fixes.sh.
- Note the skill claims 'zero-dependency' and lists no env vars but the code expects many envs and has non-trivial dependencies — that mismatch is suspicious.
- The SKILL.md contains prompt-injection patterns; avoid granting autonomous invocation until you verify those sections are harmless.
- If you want to test, run it inside an isolated container or VM with no network access and with non-sensitive HOME to see what files it creates and what processes it spawns.
- If you decide to proceed in any environment with real data, audit for network calls, review plugin autoReload behavior, and run npm audit / static analysis on the included code.
If you provide the upstream repository URL or want, I can list exact files/lines that reference network endpoints, env vars, hooks, and the prompt-injection fragments for a targeted audit.
Capability Analysis
Type: OpenClaw Skill
Name: unified-memory
Version: 5.3.0
The bundle describes an extensively complex 'Unified Memory' system with 79+ tools, including high-risk capabilities such as cloud backups (memory_cloud_backup), Git integration (memory_git_notes), and the execution of local shell scripts for 'atomic fixes' (deploy-atomic-fixes.sh). While the documentation is detailed, the actual implementation logic for these tools is missing from the provided files, and the package.json includes an 'install.sh' script which is a high-risk pattern for supply chain attacks. The broad permissions requested for the filesystem and environment variables, combined with inconsistent versioning across files (5.3.0 vs 5.2.4), suggest a high attack surface that cannot be fully verified as benign without the missing source code.
Capability Tags
Capability Assessment
Purpose & Capability
The package description claims '零依赖' (zero-dependency) and the registry lists no required env vars or binaries, but package.json declares multiple heavy dependencies (better-sqlite3, @lancedb/lancedb, chromadb, express, etc.) and included source files (~3MB merged-src.js). The SKILL.md and README describe node/mcp runtime, hooks, and scripts — so the metadata (no deps / instruction-only) does not match the actual code footprint.
Instruction Scope
SKILL.md and merged docs/included source instruct reading/writing user home paths (~/.unified-memory and ~/.openclaw/workspace), running local deploy/verify scripts, and reference environment variables (OLLAMA_URL, UNIFIED_MEMORY_*). The SKILL.md contained prompt-injection patterns (ignore-previous-instructions, system-prompt-override). The runtime instructions therefore ask the agent to access files, use env vars, and run scripts not declared in the skill metadata — scope exceeds what was advertised.
Install Mechanism
No external download/install spec is declared (instruction-only), but the skill bundle includes large code files and package.json. There is no remote URL installer (good), but installing the skill will place and potentially execute ~3MB of merged JS and scripts on disk. That is a non-trivial footprint compared to an 'instruction-only / zero-dep' claim.
Credentials
The registry lists no required env vars/credentials, but merged-docs and source reference many env vars (OLLAMA_HOST/URL, UNIFIED_MEMORY_*, paths under HOME) and integrations (embedding provider, vector DB backends). Requiring none in metadata while the code expects them is inconsistent and could lead to unexpected network access or credential usage if configured later.
Persistence & Privilege
always:false (good) but package.json declares OpenClaw hooks (before_prompt_build, agent_end) and the code writes persistent files under the user's home and a plugin directory with autoReload. That provides recurring runtime opportunities (hooks/plugins) which is reasonable for a memory plugin but increases persistence/attack surface — particularly combined with undeclared env and prompt-injection content.
How to Use
- Make sure OpenClaw is installed (local or Docker)
- Run the install command in chat:
/install unified-memory - After installation, invoke the skill by name or use
/unified-memory - Provide required inputs per the skill's parameter spec and get structured output
Version History
v5.3.0
Unified Memory v5.3.0 - Bug fixes and performance improvements
v5.2.7
v5.2.7 - 超精简发布：401个核心文件，4.3MB。文档体系重构完成：完整的中英文文档体系，涵盖快速入门、使用指南、API参考、架构设计。
v5.2.6
v5.2.6 - 文档体系重构完成：完整的中英文文档体系，涵盖快速入门、使用指南、API参考、架构设计。性能优化：混合搜索5-10倍提升，60%存储节省。原子事务管理器确保100%数据一致性。
v5.2.5
v5.2.5 - 文档体系重构：完整的中英文文档体系，涵盖快速入门、使用指南、API参考、架构设计。性能优化：混合搜索5-10倍提升，60%存储节省。原子事务管理器确保100%数据一致性。
v5.2.10
unified-memory v5.2.10
- Added MCP-CONFIG-GUIDE.md for configuration instructions.
- Added mcp_server.py, enabling new server or backend capabilities.
- No breaking changes to existing features.
v5.2.4
- Introduced a new atomic transaction manager to ensure strict data consistency for JSON and vector storage, using a two-phase commit protocol.
- Enforced durable disk writes with fsync, guaranteeing zero data loss on crash or power failure.
- Fixed LanceDB WHERE clause bugs by adding optimized in-memory filtering, resulting in 5–10× faster query performance.
- Added full support for the ChromaDB backend as an alternative to LanceDB.
- Provided one-click scripts for deploying and verifying atomic write fixes.
- Major codebase reorganization, with new config, plugin, and deployment scripts; many legacy files removed and replaced.
v5.1.0
**Unified Memory v5.1.0 — Major dream memory system refactor and optimization**
- Dream memory architecture introduced, reorganizing unstructured files into a 7-layer directory with 1760 enhanced memories and hybrid search.
- Significant performance gains: retrieval speed up 5–10×, storage use down 60%, expanded entity/tag indexing, supports 5 query types.
- New "dream" query tools and scripts for stats, tag-based queries, and full-text search.
- Extensive codebase expansion with new `memory-optimization/` module, including APIs, monitoring, sync, deduplication, and deployment scripts.
- Continued improvements based on OpenViking architecture: layered memory, type registry, async queues, smart deduplication, advanced search, lifecycle management, and multi-layer compression.
v5.0.2
集成 Claude-Mem 风格功能：记忆编辑、隐私控制、质量评估、记忆索引
v5.0.1
v5.0.1 - OpenViking unified knowledge management system with bug fixes
v5.0.0
Unified-memory v5.0.0 is a major upgrade inspired by the OpenViking architecture, transforming it from a pure storage solution into a multi-layered knowledge management system.
- Introduced a layered memory architecture (L0→L3) with structured pipelines and memory types (facts, patterns, skills, cases, events, preferences).
- Added asynchronous processing queues (embedding, semantic analysis, deduplication, archiving, indexing) for scalable background memory operations.
- Integrated a smart deduplication system for semantic merge and duplicate detection across memories.
- Enhanced memory system auto-detects types, routes, deduplicates, and processes memories with improved recall and context relevance.
- Implemented memory compressors and layered compression, saving up to 83% tokens with adaptive memory loading.
- Lifecycle management now includes auto-archiving, type-based retention, expiry, and priority protection.
- Extensive new documentation and usage examples provided (INSTALL.md, QUICKSTART_V5.md, detailed architecture docs).
v4.4.0
v4.4.0: Benchmark system, configurable Knowledge Graph, full plugin ecosystem with kg-enrich/dedup/revision plugins
v4.0.6
unified-memory v4.0.6
- Updated CHANGELOG.md to reflect latest changes.
- Bumped package version to 4.0.6 in package.json.
- No changes to functionality or APIs.
v4.0.5
v4.0.5: 重构文档系统并添加自动构建 Hook 机制
v4.0.4
fix: 5 critical async bug fixes (cloud_backup require bug, identity_tools, multimodal, search, reconsolidation)
v4.0.3
fix: make tier_tools async, await getAllMemories in memory_tier_stats
v4.0.2
fix: additional async getAllMemories bugs in MCP tool handlers (dedup, decay, memory_compose, sync)
v4.0.1
- Introduced a new HTTP server (`src/v4/http_server.js`) for v4 operations.
- Updated core modules to support HTTP-based interaction for v4 tools.
- Various internal code improvements and file refactoring in BM25, storage, semantic deduplication, and reranking logic.
- Improved compatibility and maintainability of the v4 storage and retrieval logic.
- No breaking changes to v3/v4 tool APIs.
v4.0.0
unified-memory 4.0.0
- Introduces v4.0 tools built on the new SQLite-first StorageGateway with incremental indexes.
- Adds multi-tenant team space support, including strict isolation with the new `memory_v4_team_search` tool.
- Implements distributed rate limiting, evidence TTL and revision limits, and improved WAL management.
- v4.0 tools are now recommended for all new development; legacy v3.x tools are retained for compatibility.
- Documentation updated to reflect new tools, phased enhancements, and usage patterns.
v3.8.11
Unified-memory 3.8.11 introduces v4.0 StorageGateway tools, focusing on incremental indexing, team spaces, and distributed rate limiting.
- Added v4.0 API with 17 tools for StorageGateway: includes incremental BM25 search, hybrid BM25+vector search, team space management, and evidence/revision lifecycle controls.
- v4.0 tools now support multi-tenant teams, distributed rate limiting, and transactional WAL operations.
- Legacy v3.x (79 tools) retained for backward compatibility; v4.0 strongly recommended for new development.
- Documentation updated to distinguish v4 and v3 APIs, with clear tool categorization and usage notes.
v3.8.8
unified-memory 3.8.8
- Updated core code and configuration files.
- No user-facing feature changes or breaking changes.
- Maintenance update to keep skill files in sync.
Metadata
Frequently Asked Questions
What is Unified Memory V5?
统一记忆系统 - AI Agent 专用记忆系统，支持 Context Tree、智能摘要、知识图谱、工作流引擎。零依赖，完整对标 QMD/MetaGPT. It is an AI Agent Skill for Claude Code / OpenClaw, with 774 downloads so far.
How do I install Unified Memory V5?
Run "/install unified-memory" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.
Is Unified Memory V5 free?
Yes, Unified Memory V5 is completely free, licensed under MIT-0. You can download, install and use it at no cost.
Which platforms does Unified Memory V5 support?
Unified Memory V5 is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).
Who created Unified Memory V5?
It is built and maintained by 程序员小刘 (@mouxangithub); the current version is v5.3.0.
More Skills