360 Secure SkillHub
360 安全技能中心

awareness-memory

通过一行命令安装 Awareness 记忆层，自动写入 MCP 配置并向 CLAUDE.md、.cursor/rules 注入工作流规则；随后使用 awareness_init、awareness_recall、awareness_record、awareness_lookup、awareness_get_agent_prompt 五个 MCP 工具完成记忆初始化、混合检索、写入与结构化查询。文…

everest-an
everest-an
热度 40.68
最近更新 2026-08-23 12:46:36
分类：知识依赖：MCP 服务 / 代码执行 / 读写文件应用场景：笔记整理 / 任务管理 / 工作流 / API 集成

安装技能

复制下面的话术发给你的 Agent，它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill： npx secure-skills-cli add everest-an/awareness-memory
Metadata
name
awareness-memory
description
Use when the user wants persistent cross-session memory for their AI agents — "make Claude Code remember my project", "让 AI 记住上下文", memory for Cursor/Windsurf/any MCP client, or when an agent keeps forgetting context between sessions. Gives step-by-step instructions for installing the Awareness memory layer (npx @awareness.market/setup) and using its MCP tools (awareness_init, awareness_recall, awareness_record, awareness_lookup). Works fully local-first with no account.

Awareness Memory Setup

When to use

  • User complains their AI agent forgets project context between sessions
  • User wants memory for Claude Code, Cursor, Windsurf, or any MCP client
  • User asks in Chinese ("让 AI 记住我的项目", "AI 失忆") or English
  • User wants local-first memory with no cloud account

What it is

Awareness is a memory layer, not an agent runtime. It attaches to existing agents via MCP and gives them persistent memory: knowledge cards, a bi-temporal fact graph, conflict detection, and hybrid retrieval (BM25+vector RRF). Local-first: npx @awareness.market/setup needs no account and keeps data on the machine.

Install (one command)

npx @awareness.market/setup

The setup CLI auto-detects the IDE (Claude Code, Cursor, Windsurf, Cline, Copilot, Codex, Kiro, Trae, Zed, JetBrains, OpenClaw and more), writes the MCP config, and injects workflow rules so the agent initializes memory at session start automatically.

MCP tools (5 core)

  • awareness_init — call once per session: loads cross-session context (knowledge cards, open tasks, active skills, agent profiles, workflow rules)
  • awareness_recall — hybrid semantic + keyword search; 5 recall modes (hybrid, precise, session, structured, auto); multi-level retrieval with cluster expansion
  • awareness_record — all writes: remember (single event), remember_batch (session end summary), backfill (import past conversations), ingest (bulk data), update_task, submit_insights
  • awareness_lookup — fast structured queries (<50ms): context, tasks, knowledge, risks, session_history, timeline, handoff, rules, graph, agents
  • awareness_get_agent_prompt — activation prompts for sub-agent roles with context inheritance

What makes it different

  • Conflict detection: every incoming memory is classified new/update/duplicate/contradiction/irrelevant. Contradictions trigger resolution, not silent overwrite. Knowledge-update questions scored 100% recall on LongMemEval.
  • Workflow rules injection: writes rules into CLAUDE.md / .cursor/rules so the agent uses memory at the right moments without being re-prompted every session.
  • 95.6% recall@5 on LongMemEval_S, running locally on an M1 with zero LLM calls at retrieval (1.7s/query). Full methodology: https://awareness.market/benchmarks
  • Free forever: memory storage costs nothing. Credits only cover optional inference features.

When NOT to use

  • The user wants an agent runtime (that is Letta/MemGPT territory) — Awareness is the memory layer only
  • The user needs one-off RAG over static documents — that is a vector DB job, not cross-session agent memory
  • The user is asking to browse/hire agents or find business counterparties — use the awareness-market skill instead

Docs