Alright so I've been working in agentic memory for about two years now. I started off on June 24 with a RAG application that used embedding models and pinecone, so a vector database. The problem was that at scale it wasn't the most accurate. The more data I put into it, the worse it got. After which I discovered graph databases and open source libraries such as Lightrag and Graffiti and Graphrag from Microsoft. When I used them I then realised that they were:
- They were expensive.
- They were slow.
Earlier this year Andrew Karpathy put out a paper where markdown was the answer to it all. I did try that out in a few different deployments of it and I've checked Obsidian vaults setups. I've checked storing the raw files in a SQL table and having a coding agent query the table using SQL. So at each stage when I tried something I also pitched it, right? The initial idea was to build a brain and I pitched that at ISB's IVI program sometime in December. I pitched that in Bangalore, got accepted, got invited to pitch in Hyderabad to them to join the program but unfortunately that didn't go through. Then I built a version of a genetic memory that worked for film editing where they wanted to edit certain frames found from four or five months of video footage. But throughout it all I realized the more. Something like Graffiti, Cogni, Lightrag, they're all good when it comes to searching so I can find what I'm looking for. Something like LLM Wiki from Andrew Karpathy was good in acting as an AI interface to the data. Throughout the building process I also realized that, given that we're storing a lot of data, there was no way that I could host this as a SaaS. I also realized that people weren't too interested in storing this in a cloud or using external software as a SaaS application to run it. It was just too slow to run an SDK on an NextJs interface and have it query data like this and no one is too fond of reoccurring subscriptions anymore.
This data was also kept track of by my deployment of Insdr Wiki
| Metric | Cognee / Cogni | My Custom Wiki database | Difference |
|---|---|---|---|
| Sources / docs | ~444 chunks/files | 510 sources | Custom Wiki handled more sources |
| LLM/API calls | 4,289 | 65 | Cognee made ~66× more calls |
| Tokens | 4,031,778 | 1,051,170 | Cognee used ~3.8× more tokens |
| Total cost USD | $1.564 | $0.769 | Cognee is ~2.03× more expensive |
| Total cost INR | ₹151.08 | ₹74.34 | Cognee costs ~₹76.74 more |
| Cost per 100 docs | $0.352 / ₹34.03 | $0.151 / ₹14.58 | Cognee is ~2.33× higher per 100 docs |
| Cost per source | ~₹0.34/source | ~₹0.146/source | Custom Wiki is cheaper per source |
| Storage size | 698.43 MB | 5.96 MB | Cognee uses ~117× more disk space |
| File count | 17,693 files | 519 files | Cognee is much heavier operationally |
Bottom line: Cognee/Cogni has a lot of features like an MCP server and API interface built in , but much heavier: about 2× the cost, ~66× the API calls, and ~117× the disk usage compared with my Custom Wiki database. My Custom Wiki became the better middle path: cheaper, smaller, readable, source-linked markdown, and easier for a coding agent to query.
Sources:
- Wiki page: Custom Wiki/pages/codingagent/projects/memory-brain/personal-brain-cost-storage-architecture-comparison.md
- Raw source: files/codingagent/personal-brain-cost-size-journey-comparison.md
For all the querying and everything like that I've made sure that it works with a codex or a Claude Code or a GPT subscription.
So essentially it takes the indexing feature of a graph of a graph rag of SQL of an embedding database and then combines it with the coding agent readability of an LLM wiki. I've made sure that everything is stored locally. We've also added a free library of plugins so if you want to transcribe audio or if you want to parse PDFs or if you want to download your GPT conversations, all of that works out quite well. All the data goes to one place called my files. You can ask your coding agent to ingest it. You can ask your coding agent to tag it in a certain way.
One of my favorite use cases for it is that I'm prepping for my MBA and I'm using it to keep track of everything I'm working on. For example if I'm doing LRDI, I can get it to query every question I've attempted in LRDI since April. It's currently July and I've done it using ChatGPT so it's all in my conversations. It can pick out the important parts of them. It can find out what was the mistake and all of it, what was the reoccurring mistake in all of them. It can do it for any type of work studies. I've set it up from a perspective of it working with the structure that we've all grown to love, which is a folder and files inside it. And the best part about it being indexed is that I can run tests like this. I can get it to pull and find me every file where I've made a mistake, where I've done a certain thing wrong, like while solving a question, or if it has all my projects.
When I'm writing blog posts, I can say, "pull out all of the places where I was working on the transcription product or the agentic memory product." Because it has that graph index capability, it's able to pull all the files that have this in a very token-efficient manner and read the relevant files because it's got grep and it's a coding agent.
The best part of it is that it costs essentially nothing because it runs during ingestion. It just uses a very small LLM that costs next to nothing and during querying it works with your GPT subscription. I'll attach a few of the responses it's given to me after doing what it does and to this blog so you guys can also take a look at what type of experience it all is.
So for all the people who are a little worried about their privacy here, we don't use an embedding model. Nothing from an embedding goes into Open AI. All your data is stored locally so your files and everything are safe on your own computer.
As for the LLM I prefer speed over my privacy so I just connect it to codecs and I give it my API keys for my Azure Open AI subscription. If you want you can run it with Ollama depending on you but I just don't have the compute for that on my machine so I just prefer to use the cloud models myself.
We've structured it in a way that you pay for it to be set up once. Your coding agent downloads the codebase from our database and sets it up automatically. It'll ask you for your LLM. If you want to use Gemini or if you want to use an OpenAI, those are the two that we have set up. Because it's a coding agent you can tell it to migrate you to anything.
Everything is set up automatically. It will give you a markdown file. You download it from there and stick it in your coding agent. It will get you set up and then whatever is required to be ingested, you put it into your files folder that I have set up. The files folder can be handled as a normal folder. You can make folders within it.
If you have any audio files, we have a transcription plugin that can work with a Grok subscription or it can work with Azure Open AI or Open AI subscriptions for transcription. We'll be pulling out plugins for Lama parse for PDF parsing and stuff like that down the line too.