Awareness Local: local-first memory for AI coding agents (96% R 5)

Awareness Local, a local-first MCP memory server for AI coding agents, achieves a 96% Recall@5 on the LongMemEval benchmark (ICLR 2025) with 500 human-curated questions, running fully offline on an Apple M1 with 8GB RAM and zero LLM calls. The tool provides persistent cross-session memory for Cursor, Claude Code, Copilot, Cline, and other MCP-compatible IDEs, storing memory as Markdown and indexing with SQLite FTS5 and embeddings.

Languages: English | 简体中文 /everest-an/Awareness-Market/blob/main/README.zh-CN.md Give your AI agent persistent memory. One command. No account. Works offline. Awareness Local is a local-first MCP memory server for AI coding agents. It gives Cursor, Claude Code, Copilot, Cline, and other MCP IDEs persistent memory, hybrid semantic + keyword retrieval, and reusable knowledge cards for long-running software projects. It runs a lightweight daemon on your machine, stores memory as Markdown, indexes recall with SQLite FTS5 + embeddings, and keeps your AI workflow fast, explainable, and offline-ready. npx @awareness.market/setup That's it. Your AI agent now remembers everything across sessions. AI coding agents lose context between sessions. Awareness Local provides cross-session memory recall so agents can continue work without re-explaining architecture, past decisions, pending tasks, and implementation constraints. - Persistent memory for AI coding agents - Local-first MCP server with offline support - Hybrid retrieval keyword + semantic - Knowledge card extraction for decisions, solutions, and risks npx @awareness.market/setup Then open your IDE and start coding. Awareness tools become available for recall, record, and session initialization. - Long-running codebase migrations across many sessions - Team handoffs where AI should remember prior implementation context - Personal coding workflows that need durable preferences and conventions - Multi-agent setups that share decision history and task memory Yes. Local mode works fully offline with memory stored on your machine. Memory is stored as Markdown in .awareness/ , with a local SQLite index for retrieval. No. Cloud sync is optional and can be enabled later. Any MCP-compatible IDE, including Cursor, Claude Code, Copilot, Cline, Windsurf, and others. Evaluated on LongMemEval — the industry standard benchmark for long-term conversational memory. 500 human-curated questions across 5 core capabilities. ╔══════════════════════════════════════════════════════════════╗ ║ ║ ║ Awareness Memory — LongMemEval Benchmark Results ║ ║ ───────────────────────────────────────────────── ║ ║ ║ ║ Benchmark: LongMemEval ICLR 2025 ║ ║ Dataset: 500 human-curated questions ║ ║ Variant: LongMemEval S ~115k tokens per question ║ ║ ║ ║ ┌─────────────────────────────────────────────────┐ ║ ║ │ │ ║ ║ │ Recall@1 80.2% 401 / 500 │ ║ ║ │ Recall@3 92.8% 464 / 500 │ ║ ║ │ Recall@5 96.0% 480 / 500 ◀ PRIMARY │ ║ ║ │ Recall@10 98.6% 493 / 500 │ ║ ║ │ │ ║ ║ └─────────────────────────────────────────────────┘ ║ ║ ║ ║ Method: Hybrid RRF BM25 + vector, daemon pipeline ║ ║ Embedding: multilingual-e5-small production model ║ ║ LLM Calls: 0 pure retrieval, no generation cost ║ ║ Hardware: Apple M1, 8GB RAM — 35 min total ║ ║ ║ ╚══════════════════════════════════════════════════════════════╝ ┌─────────────────────────────────────────────────────────────┐ │ Long-Term Memory Retrieval — R@5 Leaderboard │ │ LongMemEval ICLR 2025, 500 questions │ ├─────────────────────────────────┬───────────┬───────────────┤ │ System │ R@5 │ Note │ ├─────────────────────────────────┼───────────┼───────────────┤ │ MemPalace ChromaDB raw │ 96.6% │ R@5 only │ │ ★ Awareness Memory Hybrid │ 96.0% │ Hybrid RRF │ │ OMEGA │ 95.4% │ QA Accuracy │ │ Mastra GPT-5-mini │ 94.9% │ QA Accuracy │ │ Mastra GPT-4o │ 84.2% │ QA Accuracy │ │ Supermemory │ 81.6% │ QA Accuracy │ │ Zep / Graphiti │ 71.2% │ QA Accuracy │ │ GPT-4o full context │ 60.6% │ QA Accuracy │ ├─────────────────────────────────┴───────────┴───────────────┤ │ MemPalace 96.6% is Recall@5 only, not QA Accuracy. │ │ Palace hierarchy was NOT used in the evaluation. │ └─────────────────────────────────────────────────────────────┘ ┌─────────────────────────────────────────────────────────────┐ │ Awareness Memory — R@5 by Question Type │ │ │ │ knowledge-update ███████████████████████████ 98.7% │ │ multi-session ███████████████████████████▊ 99.2%│ │ single-session-asst ███████████████████████████▌ 98.2%│ │ temporal-reasoning ██████████████████████████▏ 93.2%│ │ single-session-user ██████████████████████████ 92.9%│ │ single-session-pref █████████████████████████▎ 90.0%│ │ │ │ Overall ██████████████████████████▉ 96.0%│ │ │ │ ┌───────────────────────────────────────────────┐ │ │ │ Ablation Study │ │ │ │ ───────────────────────────────────────── │ │ │ │ Vector-only: 92.6% ▓▓▓▓▓▓▓▓▓▓▓▓▓░░░ │ │ │ │ BM25-only: 91.4% ▓▓▓▓▓▓▓▓▓▓▓▓▓░░░ │ │ │ │ Hybrid RRF: 95.6% ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░ ★ │ │ │ │ 2026-04 harness run │ │ │ │ Hybrid = +3% over any single method │ │ │ └───────────────────────────────────────────────┘ │ │ │ │ arxiv.org/abs/2410.10813 awareness.market │ └─────────────────────────────────────────────────────────────┘ Zero LLM calls on retrieval daemon path . Reproducible benchmark scripts → https://github.com/everest-an/Awareness-Market/tree/main/benchmarks/longmemeval Before: Every session starts from scratch. You re-explain the codebase, re-justify decisions, watch the agent redo work. After: Your agent says "I remember you were migrating from MySQL to PostgreSQL. Last session you completed the schema changes and had 2 TODOs remaining..." Session 1 Session 2 ┌─────────────────────────┐ ┌─────────────────────────┐ │ Agent: "What database?" │ │ Agent: "I remember we │ │ You: "PostgreSQL..." │ │ chose PostgreSQL for │ │ Agent: "What framework?"│ → │ JSON support. You had │ │ You: "FastAPI..." │ │ 2 TODOs left. Let me │ │ repeat every session │ │ continue from there." │ └─────────────────────────┘ └─────────────────────────┘ | IDE | Auto-detected | Plugin | |---|---|---| Claude Code | ✅ | awareness-memory | Cursor Windsurf OpenClaw @awareness.market/openclaw-memory Cline GitHub Copilot Codex CLI Kiro Trae Zed JetBrains Junie Augment AntiGravity Jules Your IDE / AI Agent │ │ MCP Protocol localhost:37800 ▼ ┌────────────────────────────────────┐ │ Awareness Local Daemon │ │ │ │ Markdown files → Human-readable, git-friendly │ SQLite FTS5 → Fast keyword search │ Local embedding → Semantic search optional: npm i @huggingface/transformers │ Knowledge cards → Auto-extracted decisions, solutions, risks │ Web Dashboard → http://localhost:37800/ │ │ │ Cloud sync optional │ │ → One-click device-auth │ │ → Bidirectional sync │ │ → Semantic vector search │ │ → Team collaboration │ └────────────────────────────────────┘ All memories stored as Markdown files in .awareness/ — human-readable, editable, git-friendly: .awareness/ ├── memories/ │ ├── 2026-03-22 decided-to-use-postgresql.md │ ├── 2026-03-22 fixed-auth-bug.md │ └── ... ├── knowledge/ │ ├── decisions/postgresql-over-mysql.md │ └── solutions/auth-token-refresh.md ├── tasks/ │ └── open/implement-rate-limiting.md └── index.db search index, auto-rebuilt | Tool | What it does | |---|---| awareness init | Load session context — recent knowledge, tasks, rules | awareness recall | Search memories — progressive disclosure summary → full | awareness record | Save decisions, code changes, insights — with knowledge extraction | awareness lookup | Fast lookup — tasks, knowledge cards, session history, risks | awareness get agent prompt | Get agent-specific prompts for multi-agent setups | Instead of dumping everything into context, Awareness uses a two-phase recall: Phase 1: awareness recall query, detail="summary" → Lightweight index ~80 tokens each : title + summary + score → Agent reviews and picks what's relevant Phase 2: awareness recall detail="full", ids= ... → Complete content for selected items only → No truncation, no wasted tokens Visit http://localhost:37800/ to browse memories, knowledge cards, tasks, and manage cloud sync. Connect to Awareness Cloud https://awareness.market for: - Semantic vector search 100+ languages - Cross-device real-time sync - Team collaboration - Memory marketplace npx @awareness.market/setup --cloud Or click "Connect to Cloud" in the dashboard Awareness Local is part of the Awareness ecosystem: | Package | For | Install | |---|---|---| | Local daemon + MCP server | npx @awareness.market/setup | | wrap openai / wrap anthropic interceptors | pip install awareness-memory-cloud | | wrapOpenAI / wrapAnthropic interceptors | npm i @awareness-sdk/memory-cloud | | Auto-recall + auto-capture | openclaw plugins install @awareness.market/openclaw-memory | | Skills + hooks | /plugin marketplace add everest-an/Awareness-SDK → /plugin install awareness-memory@awareness | | One-command setup for 13+ IDEs | npx @awareness.market/setup | Full SDK docs: awareness.market/docs https://awareness.market/docs - Node.js 18+ - Any MCP-compatible IDE No Python, no Docker, no cloud account needed. If Awareness Local saves you from re-explaining your codebase to your AI agent, give it a ⭐ — it helps more developers discover the project and pushes it toward GitHub Trending. MIT IDE Support: Cursor https://cursor.com , Windsurf https://codeium.com/windsurf , Trae https://www.trae.sh , Zed https://zed.dev , VS Code, JetBrains. Compatible with: OpenClaw, AutoGPT, LangChain, MetaGPT. Key Technology: OMP Open Memory Protocol , LatentMAS, Shared Thought Space, One-click Deployment. Focus: Solving AI "Lobster Memory" Long-term memory loss , Automating complex workflows, Simplifying Agent setup.