In early August 2025, I broke my ankle. Badly.
I was in the hospital for almost 3 weeks. It… sucked. So I started reading a lot about AI and thinking about things. Wondering how much of my business, especially the things I don’t like, could I outsource and automate?
I’m the marketing automation guy, why not automate everything, right?🤓
I got into HippoRAG and A-MEM— papers on graph + vector memory systems for AI. It’s pretty cool. The people working on this at Google, Anthropic, etc are of course neuroscientists and physicists, but nobody knows how it works, and they’ve only used these tools for two or three months. It’s almost like we’re on even footing. It’s fun.
So I started building AutoMem—a practical implementation of their ideas that anyone could run for $5/month.
Start nerd stuff. Skip to the bottom if you don’t want technical details. 🤓
Part 1: Research 💪
HippoRAG (May 2024)
The paper that got me thinking: Neurobiologically Inspired Long-Term Memory for Large Language Models
Here’s what they figured out:
Your brain’s hippocampus doesn’t just store memories—it indexes them using three mechanisms:
- Semantic encoding (what the memory means)
- Episodic linking (how memories connect to each other)
- Pattern completion (how you reconstruct full memories from fragments)
Most RAG systems only use vector search—basically semantic encoding. They can find similar things, but they can’t connect them the way your brain does.
HippoRAG proved that if you want AI memory to work like human memory, you need graphs + vectors working together. Not just one or the other.
A-MEM (February 2025)
Then I found this: Agentic Memory for LLM Agents
What they discovered:
Memory systems based on Zettelkasten principles (basically: notes that link to other notes) outperform traditional RAG by 15-30% on reasoning tasks.
The key insights:
- Memories should associate with other memories, not just sit there waiting to be retrieved
- Associations enable multi-hop reasoning (following chains of thought)
- Graph structure emerges naturally from use (no manual organization needed)
- The system gets smarter over time as connections strengthen
OK, this was getting interesting. Graph-based memory, associations, emergent structure…
HippoRAG 2 (February 2025)
From RAG to Memory: Non-Parametric Continual Learning for Large Language Models
What they proved:
Knowledge graphs + personalized PageRank enables:
- Non-parametric continual learning (no retraining needed)
- Better than fine-tuning for long-term memory
- Scales to billions of memories
- Works by ranking associations, not just storing them
This paper basically validated everything HippoRAG and A-MEM had proposed, plus showed it could scale way beyond what anyone thought possible.
The Gap
But here’s the problem I kept running into:
Every paper ended the same way—with “future work” and “further research needed.” No one had actually built this stuff for production use.
What was missing:
- No production implementations (just research prototypes that barely ran)
- No integration with existing AI tools (good luck using it with Claude or Cursor)
- No cost-effective hosting (the examples assumed you had a research lab budget)
- No docs for normal developers (everything assumed you had a PhD in ML)
Part 2: Not much else to do 🤕
The Hospital Stay
I’d been using Claude and Cursor for years. So much time spent customizing, adding rules, and tweaking—more time configuring the interaction than actually getting productive work out of it. Every time I switched tools, the context evaporated.
After thinking about memory for so long, I began to wonder if I was approaching it the wrong way.
Timeline
August 2025 (Weeks 1-4):
- Week 1-2: Hospital stay, reading research, sketching architecture
- Week 3: First prototype—Python + SQLite
- Week 4: MCP server implementation, Docker containerization
September 2025 (Weeks 5-8):
- Week 5: Railway deployment, first memories stored
- Week 6: Full rewrite: Flask + FalkorDB + Qdrant
- Week 7: One-click setup for Claude Code and Cursor
- Week 8: Graph associations working, production-ready
Total Time: 8 weeks from concept to production
Total Cost: $5/month hosting + ~$0.10/month embeddings
Part 3: What We Built
The Architecture
Graph Database (FalkorDB):
- Memories as nodes
- Associations as edges with 11 relationship types
- Graph traversal for context building
- Relationship strength scoring (0.0-1.0)
Vector Database (Qdrant):
- OpenAI text-embedding-3-small
- Semantic similarity search
- Hybrid search with metadata filtering
6-Dimensional Hybrid Search:
- Vector similarity (semantic matching)
- Keyword matching (exact phrases)
- Tag filtering (project/topic context)
- Time decay (recent memories weighted higher)
- Importance scoring (critical info prioritized)
- Confidence weighting (validated memories rank higher)
This directly implements the HippoRAG and A-MEM architectures—but in production, at scale, for $5/month.
AutoMem Dreams at Night
OK, this is one of my favorite features and I need to tell you about it.
One of the most human-like features of AutoMem is its consolidation cycles—what we call “dreaming.” Every 6 hours, while you’re away from your keyboard, AutoMem enters a dream state. It replays the day’s memories, finding patterns you didn’t see.
Similar memories get clustered together. Weak associations fade away. Important insights get reinforced.
It’s inspired by Sharp Wave-Ripple consolidation—the same process your brain uses during sleep to strengthen memories and discover connections.
When you wake up and start coding again, AutoMem has already organized what you learned yesterday, connected it to last week’s bug fix, and figured out which architectural decisions actually mattered.
It’s not just storing memories—it’s learning from them, the same way you do when you sleep on a problem and wake up with the answer.
Pretty cool, right? 🧠✨
Making It Practical
The research papers didn’t tell you:
- How to host it affordably (Railway + Qdrant free tier = $5/mo)
- How to integrate with AI tools (Model Context Protocol)
- How to set it up in one command (
npx @verygoodplugins/mcp-automem cursor) - How to handle edge cases (embedding failures, graph cycles, memory consolidation)
That’s the gap between research and production.
Part 4: Why This Actually Matters
Academia proved the theory. I built something.
The gap between research and production typically takes 3-5 years. HippoRAG was published in February 2025. We had AutoMem in production by October. That’s fast.
Here’s what made it possible:
- Cloud hosting (Railway, Qdrant) makes infrastructure trivial
- Open-source databases are production-ready
- APIs for embeddings (OpenAI) eliminate ML ops
- Standards like MCP enable easy integration
The Problem with Everyone Else
Most AI memory systems are vector-only: Mem0, ChromaDB, most LangChain implementations, Pinecone, Weaviate.
They can’t follow reasoning chains. Can’t build context over time. No associations. It’s like… having Google search but not Wikipedia links. You can find stuff, but you can’t understand how it connects.
What Makes AutoMem Different
- Research-backed (HippoRAG, A-MEM, validated by HippoRAG 2)
- Graph + Vector (not just vector)
- 6-dimensional search (not just semantic)
- Self-organizing (associations emerge naturally, no manual tagging)
- Affordable ($5/month unlimited users vs $50-200/user/month for competitors)
- 90%+ precision on memory recall vs 60-70% for vector-only systems
We didn’t invent graph + vector memory. HippoRAG and A-MEM did the hard theoretical work.
We just made it something anyone could actually use. And in the spirit of open source, it’s completely free 🎁
Part 5: What’s Next
Q4 2025 Roadmap:
- Cloudflare Code Mode – 3-6x speed improvement (100ms → 15-30ms)
- MELODI compression – 8x memory efficiency
- Enhanced Sharp Wave-Ripple consolidation
Q1 2026 Vision:
- Multi-modal support (images, audio, documents)
- Collaborative memory (shared knowledge graphs for teams)
- Mobile apps
Long-term: AI agents that remember like humans—build context over years, follow reasoning chains, self-organize knowledge, never forget what matters, learn continuously without retraining.
Scale to billions of memories. Stay under $10/month. Keep it open source.
TLDR:
In August 2025, I broke my ankle badly enough to spend three weeks in a hospital bed reading research papers.
Eight weeks later, I had a sort of production system:
- ✅ Implements HippoRAG’s graph + vector architecture
- ✅ Implements A-MEM’s associative memory principles
- ✅ Validated by HippoRAG 2 (published one month into our build)
- ✅ Production-ready, $5/month, open source
Sometimes the best work happens when you eliminate everything else. Sometimes constraints force clarity.
And sometimes a broken ankle gives you exactly the time you need to build something that should have taken years.
Research → Reality in 8 weeks. 🚀
Try It Yourself
AutoMem is open source and production-ready. It’s completely free. Take it 🎁
Install (one command):
bash
npx @verygoodplugins/mcp-automem cursorCost: $5/month (Railway + Qdrant free tier)
Resources:
- Site https://automem.ai
- Repo: github.com/verygoodplugins/automem
- Docs: automem.ai/docs
- Quickstart: automem.ai/docs/quickstart
Questions? Want to collaborate on practical AI implementations?
Reach out: @jjack_arturo
— Jack
Research Papers Referenced
- HippoRAG (May 2024): Neurobiologically Inspired Long-Term Memory for Large Language Models
- A-MEM (February 2025): Agentic Memory for LLM Agents
- HippoRAG 2 (February 2025): From RAG to Memory: Non-Parametric Continual Learning for Large Language Models
- MELODI (October 2024): Exploring Memory Compression for Long Contexts
Technologies Used
- Graph Database: FalkorDB
- Vector Database: Qdrant
- Embeddings: OpenAI text-embedding-3-small
- Hosting: Railway
- Protocol: Model Context Protocol (MCP)
- Languages: Python, TypeScript
- Frameworks: Flask, Docker
- IDEs: Claude Code, Cursor