Openclaw Cortex Memory

OpenClaw 长期记忆插件，专为 OpenClaw AI 助手设计的智能记忆系统。

面向 OpenClaw 的长期记忆插件，集成多路检索、事件归档、图谱关系、向量化与反衰减排序，支持历史增量导入、规则反思和可观测诊断，可以将会话中零碎的记忆去噪整理、分层存储，支持对话中自动搜索回忆，吸收了 LLM wiki 的概念，帮助 Agent 在跨会话中持续积累并稳定调用高价值记忆。

核心能力

1) 语义检索

• 多路召回：keyword / BM25 / vector / graph 混合检索
• 排序融合：加权打分 + RRF + 可选 reranker
• 结果融合：可选 readFusion，支持权威融合返回
• 时序建模：memoryDecay + 命中反衰减（anti-decay）

2) 事件存储

• 分层写入：active（会话）与 archive（结构化事件）
• 摘要优先：归档记录保留 summary 与 source_text
• 向量分块：支持 summary/evidence 双通道向量记录
• 增量同步：按状态文件增量导入历史会话

3) 图谱关系

• 图谱独立层：graph/memory.jsonl 独立于 archive 文本层
• 关系追溯：每条关系可追溯 source_event_id
• 关系查询：query_graph 支持方向、关系类型、路径搜索
• 冲突治理：单值事实冲突进入队列，支持人工 accept/reject 闭环
• 可视化导出：export_graph_view 输出状态化图谱快照（含来源证据）
• 质量门禁：graphQualityMode 支持 off/warn/strict

4) 规则演进

• 反思沉淀：reflect_memory 将事件抽象为规则
• 去重治理：规则与事件均有去重控制，避免污染
• 规则复用：规则写入 CORTEX_RULES.md 供后续任务复用

5) 运维诊断

• cortex_diagnostics：模型连通、层级状态、字段对齐检查
• backfill_embeddings：支持 incremental / vector_only / full
• lint_memory_wiki：Wiki/图谱一致性巡检与修复建议
• 完整状态文件：便于快速定位同步、回填、质量问题

name: cortex-memory description: Independent skill for Cortex Memory operations in OpenClaw. Use when users ask for cross-session memory continuity, prior decisions, preferences, relationship tracing, or memory maintenance. If the memory plugin is not installed, guide users to install and enable openclaw-cortex-memory first, then continue with normal memory workflows. homepage: https://github.com/deki18/openclaw-cortex-memory metadata: { "openclaw": { "os": ["darwin", "linux", "win32"], "primaryEnv": "EMBEDDING_API_KEY" } }

Use this runtime flow:

1. Check whether plugin openclaw-cortex-memory is available.
2. If missing or disabled, run the install bootstrap guide in {baseDir}/references/agent-manual.md with README-consistent install order.
3. After plugin is enabled, run memory retrieval or write operations.

After plugin is ready, use this default workflow:

1. Retrieve candidate memory with search_memory.
2. If the user asks about entity links or dependency paths, use query_graph.
3. If query_graph returns conflict_hint, use list_graph_conflicts and ask user confirmation before resolve_graph_conflict.
4. For graph status snapshots or projection consistency checks, use export_graph_view and lint_memory_wiki.
5. Answer with retrieved evidence first, then add reasoning.
6. If the turn introduces durable new information, persist it with store_event.
7. For maintenance windows or session wrap-up, run sync_memory and optionally reflect_memory.

If tool calls fail:

1. Check plugin availability first (openclaw plugins list / openclaw plugins inspect openclaw-cortex-memory).
2. If visible, run cortex_diagnostics (or legacy alias diagnostics).
3. Tell the user the plugin is unavailable or misconfigured.
4. Continue with a normal answer without fabricating memory hits.

Do not use this skill for:

• Real-time web facts (weather, stock, breaking news).
• One-off temporary context that should not be persisted.

Use these local references when needed:

• Install and bootstrap playbook: {baseDir}/references/agent-manual.md
• System prompt template: {baseDir}/references/system-prompt-template.md
• Tool details: {baseDir}/references/tools.md
• Config and dependency setup: {baseDir}/references/configuration.md
• Publish and validation checklist: {baseDir}/references/publish-checklist.md

For environment preflight, use {baseDir}/references/publish-checklist.md.