Everyone is racing to give models a memory. Almost no one agrees on what it should be allowed to keep.
Six camps are pulling the word "memory" in different directions — from test-time architectures to the right to be forgotten. This plate ranks what's actually load-bearing right now, and where the argument is heating up fastest.
Read byFrontier-lab research leadsandHyperscaler platform CTOsandResponsible-AI leadership
The six camps, ranked by what's load-bearing right nowauthority mass = Σ(source weights) · recomposes live as you filter or the agent learns
The graph beneath — every one of 24 sources, weighted & linked, the substrate this infographic is composed from
Showing 24 / 24 sources · 6 facets
Weighting legend
lower weight
corroborated, cited
load-bearing consensus
weight = authority × recency × corroboration. Node size and ring glow track the score; faint links show which source leans on which.
Source key
arXiv / lab paper
podcast / talk
X / Bluesky voice
Byron's own note
Facets click to isolate
What changed overnight
New
Titans: Learning to Memorize at Test Time indexed — corroborated by 4 later papers, entered Architecture at high weight.
Trend
Memory Safety & Privacy is now the fastest-rising facet: share of new sources up 3.1× since April.
New voice
Abeba Birhane added to the Safety cluster — dataset-memorization audit thread, corroborates Mitchell.
Reweight
Biological Analogy down-weighted (recency decay); Unlearning benchmarks up on fresh corroboration.
Clip the plate exactly as it's filtered right now — the headline, the weighted facets, the trend — into one self-contained block. The person you send it to gets your reading of the field, not a link they'll never open.
⌘
Live filter
Talk to the plate. It re-composes — which facets stay, which sources survive the cut.
Send this block
LEGEND · block on "memory"
The state of the field
Weighted · self-improving · pass #61opens as a live block, not an image
Recipient sees your exact filter and can keep talking to the graph from where you left it.