AI Agents Shouldn't Be Strangers Every Morning

What it does

MemoMind is a local-first memory layer for Claude Code that turns conversations into a persistent knowledge graph. It runs entirely on your machine using PostgreSQL and pgvector, automatically extracting facts and relationships from chats so your agent recalls preferences, decisions, and project context across sessions. The system also imports historical conversations from ChatGPT and Gemini, plus personal activity logs from the DayLife app, merging them into a single searchable timeline.

The interesting bit

Instead of dumping a static markdown file into context, it uses a four-way hybrid retrieval pipeline—semantic search, keyword matching, knowledge graph traversal, and temporal lookup—to surface only relevant memories without wasting tokens. A reflect function lets the agent synthesize insights across the entire memory store rather than just fetching isolated facts.

Key highlights

• 100% local stack: PostgreSQL, pgvector, and GPU-accelerated embeddings via CUDA; no data leaves the machine
• Automatic extraction: a split-LLM setup uses a cheap model for fact extraction and a stronger one for consolidation and merging
• Cross-platform imports: ingests ChatGPT and Gemini histories via companion exporters, and DayLife activity timelines, with source tracing back to original conversations
• 4-way hybrid retrieval: semantic + BM25 + knowledge graph + temporal search; the author reports keyword recall in ~20–33 ms and semantic recall in ~235–430 ms in personal production use
• MCP-native: speaks the Model Context Protocol over stdio, designed specifically for MCP-based coding agents

Caveats

• No published LongMemEval or third-party benchmark scores; the performance and scale figures (50,100+ nodes, 2.9M+ links) come from the author’s own multi-year personal deployment
• GPU acceleration requires NVIDIA/CUDA; Apple Silicon and AMD GPUs are not mentioned
• Importing ChatGPT and Gemini conversations depends on separate companion repositories, not built-in scrapers

Verdict

Worth exploring if you run Claude Code locally and want a brain that doesn’t leak data or reset every morning; less compelling if you need independently validated benchmarks or non-NVIDIA GPU support.