Project · Signal & memory · July 2026

LatentFeed: a daily signal from the AI firehose

Keeping up with AI research is itself a long-horizon problem. The field publishes faster than any human can read, across venues that don't talk to each other — papers on arXiv, lab blogs, trending repositories, forum threads. Attention is the scarcest resource a researcher has, and most of it gets spent on triage.

LatentFeed automates the triage. Every morning it sweeps six sources — arXiv, frontier-lab blogs, GitHub Trending, Hacker News, Papers with Code, and Reddit — deduplicates and filters for relevance, summarizes each item with an LLM, sorts everything into ten curated research topics, and delivers a single clean briefing: injected straight into your Obsidian daily note, emailed to subscribers, or both. By the time you sit down with coffee, the day's signal is already separated from the noise.

How it works

From six firehoses to one briefing

Runs unattended every morning, on a schedule

01Collect

Six pluggable sources fan out in one sweep — each independently toggleable, each failing gracefully.

02Distill

URL-level deduplication, then keyword-relevance filtering tuned to AI/ML research topics.

03Summarize

Each surviving item is summarized by an LLM — cloud API or fully local — with graceful fallback.

04Organize

Auto-categorization into ten curated topics, from frontier labs to interpretability to systems.

05Deliver

A formatted digest lands in your Obsidian daily note and inbox as styled email — idempotently, without touching your other notes.

scroll →

Small tool, deliberate engineering

LatentFeed started as an internal tool and became part of how we think about memory systems: an agent that works over weeks needs exactly this shape of machinery — continuous ingestion, aggressive filtering, and durable storage of what matters.

Want the daily briefing, or want to talk about signal pipelines? Write to research@r2m.ai.