Static embeddings offer unmatched throughput, but they also suffer from accuracy loss compared to traditional embedding models. Can we make them better for retrieval? We tried several things: ✅️ raw maxsim scoring on the per-token embeddings ✅️ training a small adapter model ✅️ changing the distillation training target and teacher While none of these saw the results we wanted, it's an excellent dive into static embeddings and how they do (and don't) work! Blog:

Recent research has highlighted the promise of scaling memory embeddings in LLM training. While Engram and STEM index memory by token identity or local n-grams, can we design more flexible memory routing that captures how each token’s meaning changes with context? This paper introduces Mixture-of-Memory Embeddings (MoME), which uses a learned router to sparsely select among multiple memory slots for each token. The architecture outperforms strong memory baselines across three model families. Interestingly, both qualitative and quantitative analyses show that the learned routers’ activations correlate with the context-dependent senses of polysemous words. The authors also train a sub-billion-parameter model that achieves competitive CORE-22 performance against similarly sized base models, including Qwen3-0.6B and Llama 3.2-1B. Code and pretrained models are publicly available!

🔎 A provocative paper: "agentic search needs no embeddings or vector DB—just grep the raw corpus directly." Title: Beyond Semantic Similarity: Rethinking Retrieval for Agentic Search via Direct Corpus Interaction URL: ❓ What is Direct Corpus Interaction (DCI)? 💡 A retrieval paradigm where the agent explores the raw corpus directly via grep, find, and shell commands—no embedding model, vector index, or retrieval API. It needs no offline indexing and adapts naturally to evolving local corpora. ❓ Why aren't conventional retrievers enough? 💡 Sparse or dense, they compress the corpus through a fixed similarity interface into top-k before reasoning begins. That makes exact lexical matching, combining weak clues, and checking local context hard—and evidence filtered out early can't be recovered no matter how strong the downstream reasoning. That's fatal for multi-step agent search. ❓ Can it really win without a retriever? 💡 Yes. On BrowseComp-Plus, swapping the retriever for DCI on the same Sonnet 4.6 lifts accuracy 69.0 → 80.0% (+11.0) while cutting cost 29.4%. Multi-hop QA averages 83.0 (+30.7 over the strongest baseline); IR ranking hits 68.5 NDCG@10 (+21.5). Even the lightweight GPT-5.4 nano version beats many baselines. ❓ What's actually driving it? 💡 The authors call it "retrieval interface resolution." Trajectory analysis shows the edge comes less from surfacing more gold documents and more from high-resolution local search and verification after reaching them. #AIAgents# #RAG#

“Budget” and “financials” are different words, but embeddings understand they’re related. That’s the foundation behind semantic search and one of the core building blocks of modern multimodal systems. Learn how embeddings power retrieval across text, audio, images, and video in Building Multimodal Data Pipelines:

Me in 2012 instead of learning: DSA System Design AI / ML Generative AI LLMs AI Agents RAG Vector Databases Embeddings Fine-tuning Prompt Engineering MCP AI Coding Agents Python Java C C++ Go Rust JavaScript TypeScript HTML CSS React Next.js Vue Angular Node.js Express.js NestJS Spring Boot Django FastAPI REST APIs GraphQL WebSockets SQL PostgreSQL MySQL MongoDB Redis Elasticsearch Kafka RabbitMQ Docker Kubernetes Terraform AWS Azure GCP Linux Git GitHub CI/CD DevOps Microservices Serverless Cloud Cybersecurity Data Engineering Data Science MLOps Data Analytics Blockchain Web3 Smart Contracts Solidity Edge Computing Distributed Systems Observability Prometheus Grafana OpenTelemetry Supabase Firebase Vercel Cloudflare PostgreSQL + pgvector LangChain LlamaIndex PyTorch TensorFlow Hugging Face Ollama OpenAI APIs Claude APIs Gemini APIs Model Context Protocol AI Infrastructure GPU Computing CUDA

the next unlock is inference-native context liquidity. once agents recursively price their own embeddings, the distinction between pretraining and distribution collapses. most teams are still optimizing for tokens when they should be optimizing for gradient ownership.

We have 85 PRs merged last week and some big announcements: - cognee 1.6.0 release now runs without an LLM API key: local GLiNER extraction + fastembed embeddings (demo mode) 🔥 - Cloud: GitHub connector landed, connect your whole org and load all your projects in minutes +  live-streamed search answers - improve() became an explicit orchestrator, and recall routing got stricter — unresolvable dataset names now fail loudly instead of being silently dropped - Typed edges for custom graph models + Presort: folder pre-organization and a JSON graph-model DSL - Security hardening: random defaults replace the hardcoded token secret and default-user password, external ref schemas rejected, MCP transport origin guards - cognee-mcp: PyPI Trusted Publishing, X-Api-Key auth for self-hosted mode, recall results now summarize memory hits

Updates: Auto-list: pick a discount once, your staking rewards list themselves on the book Partial unstake: take out any amount, the rest keeps earning More models, one key: audio and embeddings on the gateway

Conversational voice agents need to know when a different person is speaking, but not every background voice should count. We combined speaker embeddings with a post-trained audio-language model to determine when a speaker change matters.