Tiered Context Manager

多Agent协作的智能会话上下文管理系统。当需要管理AI agent的长会话压缩、多层记忆分层、跨Agent知识共享时激活。支持L1/L2/L3分层压缩、实时监控、统计分析。用于OpenClaw agent的上下文管理优化。

This package implements what it claims (tiered compression and cross-agent sharing) but contains several red flags you should consider before installing or running it: - Hard-coded absolute paths: many scripts use Windows-specific paths (e.g. E:\\zhuazhua\\...) to read/write sessions, inbox, shared memory, stats and reports. If your environment has such paths mounted those files will be read/written; if not the scripts may fail. Ask the author or replace these with configurable paths before use. - Unfiltered extraction & publishing: cross_agent_context extracts 'preferences', 'technical', 'commands', 'paths' from session content and appends them to a shared knowledge file. That can leak secrets, config values, or file paths from sessions to other agents. Do not run against sessions that may contain secrets until you audit and restrict extraction rules and output destinations. - File modification & cleanup: memory_tiering can rewrite file frontmatter, move files to an archive, or delete expired ephemeral files. Back up any workspace before running or run in an isolated sandbox. - Loading OpenClaw runtime by path: index.js uses require.resolve and then requires a specific dist filename (compact.runtime-C0J2-J-T.js). This is brittle and unusual; verify what that file is and prefer documented APIs instead of importing a specific build artifact. - Run in a sandbox first: execute the scripts in a controlled environment with non-sensitive data and override default directories to a temporary location. Search the codebase for all occurrences of absolute paths and for any file-deleting operations. If you intend to use it in production, require the author to make paths configurable and to add filtering/whitelisting of extracted content, and provide documentation about what is written to disk. What would increase confidence: a version that exposes configurable directories (no hard-coded absolute paths), documents and limits what is extracted/published, and uses documented OpenClaw APIs rather than requiring specific dist filenames. If the author provides a config/default.json that clearly maps default storage to a safe sandboxed directory, and demonstrates tests showing no sensitive data is published, the assessment could move toward benign.

Type: OpenClaw Skill Name: tiered-context-manager Version: 1.0.0 The tiered-context-manager bundle is a sophisticated system for managing AI session history through multi-level compression (L1/L2/L3) and memory tiering. While the code makes extensive use of local file system operations (reading/writing sessions and memory files) and includes hardcoded Windows-specific paths (e.g., 'E:\zhuazhua\...'), these behaviors are consistent with its stated purpose of optimizing OpenClaw agent context. There is no evidence of data exfiltration, network communication, or malicious prompt injection; the 'L3' compression logic uses the agent's own capabilities to summarize its history locally. The dynamic module loading in index.js appears to be a deep integration with the host application's internal runtime rather than an obfuscation technique.