AI 보조자를 위한 로컬 우선 MCP 메모리 시스템에 대한 피드백은 무엇입니까?

Local Memory MCP v1 is a local-first personal RAG memory system for AI assistants. It stores text chunks plus lightweight metadata in a local ChromaDB, then exposes MCP tools so a new LLM session can quickly recover user context. This project is built for technical users who want to self-host and control their own data. It is not a SaaS product. AIX (AI eXperience) means designing for how LLMs actually consume context: - Prefer clear text chunks over rigid document schemas. - Keep metadata minimal but useful: timestamps, confidence, supersedes links, deprecation flags. - Preserve history with version chains instead of destructive overwrites. - Return warning-rich tool responses so the model can self-correct write behavior. The goal is practical retrieval quality and reliable AI behavior, not perfect human taxonomies. [Assistant via MCP Client] | v [run_mcp_v1_stdio.py | run_mcp_v1_http_sse.py] | v [src/mcp_server_v1.py] / | \ v v v [src/vector_store.py] [src/reconciliation.py] [src/health_monitor.py] | | v v [Local ChromaDB] [Reconciliation Log Collection] Write path: store orupdate writes a chunk.- Reconciliation checks for overlap/conflict signals. - The system returns warnings/self-heal hints when writes look risky. Read path: search runs semantic retrieval.- Ranking blends similarity with lightweight lexical/recency signals. - Deprecated chunks stay hidden by default unless explicitly requested. Current v1 capabilities: - MCP tools for store ,search ,update ,delete ,get_chunk ,get_evolution_chain . - Versioned updates ( strategy="version" ) with supersedes chains. - Soft delete by default (history retained), optional hard delete. - Heuristic reconciliation and conflict logging. - Warning-first write responses with structured warnings[] and self-heal fields. - Health checks for oversized chunks and unresolved conflicts. - Local backup/restore commands for the persisted vector DB. - Stdio transport and SSE transport for MCP clients. - Optional auth modes for SSE: none (local-only),bearer , oroauth . Use one of the two paths below. If you are unsure, choose Path A (Docker). Best for most users. This avoids local Python/venv setup. Prerequisite: - Docker Desktop (or Docker Engine) is installed and running. - Clone the repo: git clone cd local-memory-mcp - Start the service: docker compose up --build -d - Verify endpoints: http://localhost:8000/mcp http://localhost:8000/sse http://localhost:8000/messages/ http://localhost:8000/health - Stop when finished: docker compose down For config mounts, volumes, and stdio-in-Docker details, see docs/docker.md . Use this when you want direct local Python control and stdio-first desktop workflows. Prerequisites: - Python 3.11+ pip - Windows PowerShell or a POSIX shell - Clone and install dependencies: git clone cd local-memory-mcp python -m venv .venv Windows: .\.venv\Scripts\Activate.ps1 pip install -r requirements.txt macOS/Linux: source .venv/bin/activate pip install -r requirements.txt - Ensure the embedding model is available locally (one-time): python -c "from sentence_transformers import SentenceTransformer; SentenceTransformer('all-MiniLM-L6-v2')" - Optional: create a local config override (kept out of git): Windows PowerShell: Copy-Item config.example.json config.json macOS/Linux: cp config.example.json config.json If you skip this, built-in defaults are used (local-first, MCP_AUTH_MODE=none ). - Run a direct local verification (no MCP client required yet): python examples/try_local.py This performs one write and one retrieval, and creates ./chroma_db automatically on first write. - Run MCP over stdio (recommended starting point for local runtime): python run_mcp_v1_stdio.py - Optional: run SSE server: python run_mcp_v1_http_sse.py - Optional: expose SSE through an external relay workspace: - Keep relay scripts/config in a separate folder outside this repository. - Point the relay to http://localhost:8000 . tool: store input: { "text": "Example user context: weekday focus block is 6:30-9:00 AM as the current default schedule." } tool: search input: { "query": "current deep work schedule", "top_k": 5 } Use a focused retrieval pass, then summarize: search("current work schedule and constraints") search("active priorities this month") search("current preferences and hard boundaries") Then synthesize only active, non-deprecated chunks into a short session brief for the new model instance. More sample chunks and retrieval flows are in examples/ . - Real memory data is stored in ./chroma_db by default and is generated locally at runtime. - Local DB and backup folders are ignored by git ( chroma_db/ ,backups/ , and common DB file extensions). config.json is treated as local machine config and is ignored by git.- Keep commit-safe templates in config.example.json and.env.example . - Default posture is local-first and user-controlled. - Data is stored in local ChromaDB files under the configured persist directory. - The server itself does not require a cloud backend. - O