The Problem Nobody Talks About
Every AI assistant has amnesia. Close the conversation, start a new one, and it has no idea what you discussed yesterday. The decisions you made, the bugs you fixed, the architecture you chose - gone.
People solve this in different ways:
- ●Obsidian with backlinks and graph view for manual knowledge management
- ●CPR (Compress/Preserve/Resume) scripts that dump session context to files
- ●CLAUDE.md files that give Claude persistent project context
- ●Memory systems with flat markdown files and manual indexes
We tried all of them. They all have the same limitation: retrieval is keyword-based. If you do not remember the exact word used in a note from three weeks ago, you will not find it.
What We Actually Built
We deployed LightRAG on a $7/month VPS. LightRAG is a graph-augmented retrieval system with 30,000+ stars on GitHub, published at EMNLP 2025.
Here is what it does that flat files cannot:
1. It builds a knowledge graph automatically. Feed it markdown files and it extracts entities (people, projects, tools, decisions) and their relationships. No manual backlinks needed. No wikilinks. No tagging. It reads your content and figures out how things connect.
2. It searches by meaning, not keywords. Ask "what did we decide about pricing?" and it finds the answer even if the word "pricing" never appears in the document. It might be in a note that says "set the monthly plan at $9 and annual at $79." Semantic search handles this. Keyword grep does not.
3. It combines vector search with graph traversal. Traditional RAG finds similar text chunks. LightRAG also traverses the knowledge graph to find related entities across documents. If you ask about a client project, it connects the client to their contract, their support tickets, and the architecture decisions made for their deployment.
4. It runs 24/7 on a VPS. Not on your laptop. Not in a browser tab. On a server that is always on. Claude Code queries it via MCP. A Telegram bot queries it from a phone. The knowledge base is accessible from anywhere.
Why Not Just Use Obsidian?
Obsidian is a fantastic writing tool with a massive ecosystem (2,700+ plugins, 180,000+ Discord members, and a dedicated community building incredible things). If you already use Obsidian, you can even add LightRAG to it via the Neural Composer plugin, or connect Claude Code to your vault through MCP servers.
We chose a different approach: running LightRAG standalone on a VPS.
What Obsidian gives you:
- ●Manual backlinks with
[[wikilinks]] - ●A visual graph of your note connections
- ●Community plugins for semantic search (Smart Connections, Copilot)
- ●MCP servers for Claude Code integration
- ●A great editor with local-first data ownership
What a standalone LightRAG deployment adds:
- ●Automatic entity and relationship extraction across all your documents
- ●Semantic search plus graph traversal in a single query
- ●A REST API accessible from any tool (Claude Code, Telegram, scripts, automation workflows)
- ●A WebUI with interactive knowledge graph visualization
- ●24/7 availability on a server, queryable from your phone without the desktop app running
Both approaches work. The standalone VPS approach appealed to us because we wanted a knowledge base that runs as a service, accessible from Claude Code, a Telegram bot, and automation workflows, all without needing a desktop app open.
LightRAG automatically discovers connections across documents. You feed it your notes, session logs, and project docs, and the knowledge graph builds itself. For a knowledge base that grows over time, that automatic discovery is powerful.
You can still write in Obsidian, VS Code, or any editor. LightRAG does not care what tool created the files. It just needs text.
The Architecture
The stack runs entirely on a single VPS. LightRAG handles the knowledge graph and vector database. An LLM handles entity extraction and embeddings. Claude Code and a Telegram bot query the system. The whole thing runs in Docker alongside your existing services.
The total additional cost to run this on a VPS you already have: effectively zero. The LLM and embedding APIs have generous free tiers, and LightRAG is open source under the MIT license.
What We Feed It
Get the Weekly IT + AI Roundup
What changed this week in NinjaOne, ServiceNow, CrowdStrike, and AI. One email, every Monday.
No spam, unsubscribe anytime. Privacy Policy
Not everything belongs in a knowledge graph. We are selective about what gets ingested:
What goes in:
- ●Project decision logs (why we chose X over Y)
- ●Session summaries (what was done, what was decided, what is next)
- ●Architecture notes (how services connect, what ports they use)
- ●Skill definitions (what each automation does and when to trigger it)
- ●Blog posts (content we have published)
- ●Feedback memories (user preferences, workflow corrections)
What stays out:
- ●API keys, tokens, and credentials (security risk)
- ●Raw code files (too noisy, better served by code search)
- ●Temporary debugging notes (ephemeral, not worth indexing)
- ●Anything already in git history (git log is the authoritative source)
The ingestion is incremental. New documents get pushed automatically as part of your workflow. The knowledge graph grows organically without manual maintenance.
How Queries Work
LightRAG supports multiple query modes that combine vector similarity with graph traversal. Ask a question in plain English and it finds relevant information by meaning, traces relationships across documents, and returns a synthesized answer with source attribution.
The difference from regular search: a single question can pull context from five different documents and connect them through the knowledge graph. No single file has the complete answer, but the graph assembles it.
Real Examples
Before: "What was the issue with the plugin permissions?"
Search: grep for "permission" across 30+ files. Hope the right keyword was used. Read through multiple partial matches. Piece together the story manually.
After: Same question returns a synthesized answer assembled from three different documents: a session log, a setup guide, and a configuration file. The knowledge graph connected the entities across all three and returned the full picture in one response.
Who This Is For
Developers maintaining multiple projects who need to recall past decisions without searching through hundreds of files.
IT teams managing client documentation, runbooks, and incident post-mortems that need to be searchable by meaning, not just keywords.
Solo founders building products where the context lives in one person's head and needs to be externalized for AI assistants to use.
Anyone who has thought "I know I wrote this down somewhere" and spent 20 minutes trying to find it.
The Bigger Picture
The AI memory problem is real. Every AI assistant forgets everything between conversations. The solutions so far have been band-aids: longer context windows, session dumps, manual summaries.
LightRAG is different because it does not just store text. It understands relationships. It knows that your server runs multiple services, that those services connect to each other, and that a change in one affects the others. It knows that your Telegram bot uses a skill that calls an API that returns data from a database. These connections are not in any single document. They emerge from the graph.
That is what a knowledge graph gives you that flat files cannot: emergent connections that compound over time.
Get Started
LightRAG is open source (MIT license) and runs in Docker on any Linux VPS.
Want to build this yourself? Our LightRAG Pro Track walks through deployment, ingestion strategies, Claude Code integration, phone access, and multimodal embeddings step by step.
Not sure which AI tools fit your workflow? Take the free quiz and get matched in 2 minutes.