⚡ Learn the BaseRetriever pattern
⚡ See real code snippets
⚡ Take the “Which Retriever Are You?” quiz

Read to get smarter answers? Let me know which retriever you are 🙂

dub.sh/cognee-retri...

Bottom line: if you’re building agents, assitants, or automated workflows, it’s time to evolve from “data lake” to AI memory “lake”.

- Read the deep dive ➡️ dub.sh/file-based-m...

- GitHub ➡️ github.com/topoteretes/...

- Join us on Discord ➡️ discord.com/invite/tV7pr...

We also introduce dreamify - our optimization engine that tunes chunk sizes, retriever configs & prompts in real time for max accuracy and latency ✨

Why file-based?

• Cheap, cloud-native (S3, GCS)

• Scales linearly with data growth

• Easy diff + version control

• Plays nicely with existing ETL & BI stacks

It’s a living system:

1️⃣ User adds data

2️⃣ Data is cognified

3️⃣ Search & reasoning improve

4️⃣ Feedback flows in

5️⃣ System self-optimizes

…and the loop keeps compounding value. ♻️

🔑 Key insight: Data → Memory → Intelligence

Our pipeline “cognifies” every file into graphs, giving agents memory - just like a human mind. So let’s see how 👇🏼

First, why care about AI memory?

LLMs are brilliant—until they meet your fragmented data. They forget, hallucinate, or drown in silos. File-based AI memory bridges that gap, turning raw files into contextual intelligence. 📂🧠

Explore the research: arxiv.org/abs/2505.24478

Our GitHub: github.com/topoteretes/...

Taken together, the results support the use of hyperparameter optimization as a routine part of deploying retrieval-augmented QA systems. Gains are possible and sometimes substantial, but they are also dependent on task design, metric selection, and evaluation procedure.

We evaluate on three established multi-hop QA benchmarks: HotPotQA, TwoWikiMultiHop, and Musique. Each configuration is scored using one of three metrics: exact match (EM), token-level F1, or correctness.

We present a structured study of hyperparameter
optimization in graph-based RAG systems, with a focus on tasks that combine unstructured inputs, knowledge graph construction, retrieval, and generation.

Building AI memory and data pipelines to populate them is tricky. The performance of these pipelines depends
heavily on a wide range of configuration choices, including chunk size, retriever type, top-k thresholds, and prompt templates.

Why does AI memory matter?

LLMs can’t give us details about our data, they "forget" or simply don’t know the details.

Full write-up → www.cognee.ai/blog/fundame...

If you’re exploring how to blend vectors and graphs for richer retrieval, we build exactly that at @cognee.bsky.social - DMs open for a chat!

@PGvector

If your need something cheap and a way to get started, pgvector is the key. If you need something to run in production with large volumes, well, maybe you will run into trouble there. Still, it will do a lot of heavy lifting for you