blog

Writing

Engineering posts about the AI Trio, benchmarks, and local-first AI infrastructure.

benchmarks · Sep 18, 2026

We read every memory benchmark so you don't have to: LoCoMo and LongMemEval, honestly explained

What the two standard benchmarks actually measure, why retrieval-only recall and LLM-judge accuracy are two different games, and the five methodology traps — judge generosity, top-k gaming, category exclusions — that make memory leaderboards untrustworthy.

Read the post
architecture · Sep 18, 2026

The frontier stopped deleting: ADD-only memory and the end of overwrite

In 2026 the best memory systems converged on the same design: never update, never delete — append and invalidate. Why overwrite was a mistake, how Mem0 and Graphiti got there, and what it looks like in a local SQLite file.

Read the post
retrieval · Sep 18, 2026

RAG is not memory: what the leaderboards teach about hybrid retrieval

BM25+vector beats either alone, reranking is the cheapest +3.4pp in the field, and entity linking is the underused signal. Why memory retrieval is evidence assembly, not ranking — and the local stack that does it with no API key.

Read the post
research · Sep 12, 2026

The 2026 agent-memory landscape: what we learned, and what MemOS will adopt

We read the 2025–2026 memory wave end to end — Letta's sleep-time compute, HippoRAG 2, A-Mem, MIRIX, Mem0's industry report, LongMemEval-V2, and the compression literature. Five paradigms, a benchmark crisis, and eight concrete upgrades coming to MemOS.

Read the post
open source · Aug 19, 2026

Why agent memory shouldn't cost $249 a month

Mem0's platform charges $19–$249/month and keeps your memories in their cloud. MemOS is MIT-licensed, runs on a SQLite file on your disk, and includes graph memory for free. Here's the full breakdown — pricing, features, and what 'open source' actually means for each.

Read the post
benchmarks · Aug 12, 2026

BEAM-1M results: 95.9% recall with local SQLite and Gemma-300M embeddings

How MemOS approaches long-context memory benchmarks with a fully local stack — hybrid retrieval, temporal validity, and trust scoring.

Read the post