I just pushed the current SOTA on the LoCoMo long-term memory benchmark for agents: 80.1 % accuracy using only: -BGE-large-en-v1.5 (1024d) + FAISS

-Custom “MCA” gravitational ranking (keyword coverage + importance + frequency)

-BM25 sparse retrieval

-Direct Cross-Encoder reranking (bge-reranker-v2-m3) on the full union (~120-150 docs)

-Gpt-4o-mini only for final answer generation and judging (everything else is open weights or classic)

Repo: https://github.com/vac-architector/VAC-Memory-System Key tricks that finally broke 80% :

-MCA-first filter (coverage ≥ 0.1 → top-30) — catches exact-keyword questions early

-Feeding the entire union straight into Cross-Encoder (112–135 documents) instead of pre-filtering

-Proper query instruction for BGE-large (the classic “Represent this sentence for searching relevant passages”)

The whole pipeline runs in < 3s per query on a single RTX 4090. LoCoMo is currently the hardest public long-term memory benchmark (5.880 real human–agent conversations, multi-hop, temporal, negation, etc.).

Beating Mem0 official baseline by ~12–14 pp with fully open components feels pretty good. Would love feedback, especially from people who are also grinding on agent memory systems.

My background: My path didn't start in an IT office, but in Columbus, Ohio, where I worked as a handyman after leaving my job on the cell towers. The decision came from necessity: I bought a powerful PC on installments and resolved to create something that would change my life.

I had no experience, but I had an idea. Using Claude CLI as my sole mentor, I focused on architecture, not syntax.

Over 4.5 months of work, I engineered and created the VAC Memory System. To prove its value, I tested it on the toughest RAG benchmark—LoCoMo. Today, my system shows an overall result of 80.1% and a phenomenal 87.78% in the "Commonsense" category.

This is more than just code; it is the result of faith in an idea. I showed that by using modern tools, it is possible to achieve SOTA-level performance and create serious technology, regardless of your starting point. I highly anticipate your feedback.

X-ray: a Python library for finding bad redactions in PDF documents

Linus Torvalds on Linus Torvalds (1999)

Show HN: Greed.js – Run PyTorch code in browser using WebGPU

The size of 3I/ATLAS from non-gravitational acceleration

I didn't realize my LG TV was spying on me until I turned off this setting

Polyglot-image: A disk image that boots on multiple architectures

Separating AI "context" from models so teams can switch without losing state

Ask HN: How to find summer internships at startups in SF?

Kick Assembler Manual

China just carried out its second reusable launch attempt in three weeks

Can a Transformer "Learn" Economic Relationships? Revisiting the Lucas Critique

Lsp-CLI: CLI language client for LSP language servers

Researchers develop a camera that can focus on different distances at once

Writing DataFrame-Agnostic Python Code with Narwhals – Real Python

Mostly Automated Proof Repair for Verified Libraries [pdf]

Why do students leave exams early?

Video to ASCII React Component

AI-powered mock interviews with scoring, diagnostics, and targeted drills

Conductor: Enforce a "Spec → Plan → Code" Workflow in the Gemini CLI

Expertise and Automation: Effects on Wages and Jobs (NBER Working Paper)

We removed 80% of our agent's tools

Show HN: Oblique Strategies for Claude Code

80.1 % on LoCoMo Long-Term Memory Benchmark with a pure open-source RAG pipeline

We asked four AI coding agents to rebuild Minesweeper–the results were explosive

Apple Clings to Samsung as RAM Prices Soar

The Geopolitics of Net Zero

Why Custom Evals Matter for Production LLMs

Inference Economics 101: Reserved Compute versus Inference APIs

US regulators approve Wegovy pill, first oral medication to treat obesity

Patrick Roach Sues Verizon for Refusing to Unlock His Phone (and Wins)

Ask HN: Which LLM has the best "study and learn" functionality?