A&M Records released Temple of the Dog in April 1991. It sold 70,000 copies. In September, Ten dropped. In October, Badmotorfinger. A&M reissued the album the following year — same record, new context — and it went platinum. The music hadn’t changed; who had made it now meant something different to anyone searching the back catalog.
LLM agents accumulate context as they work — every output, every intermediate result added to the window. Research from JetBrains found that this growing archive quickly becomes noise: the model’s effective context is small, so most of what’s been added is functionally unreachable. Items don’t fail — they become unreachable because no signal is specific enough to find them in the mass.
The read-later queue has the same problem, not because the saved articles are bad, but because the queue has no retrieval architecture — nothing arrives at the right moment with enough specificity to pull out a particular item. The article waits for a query that never comes in a form the queue can handle.
A dormant archive item is just waiting for a signal specific enough to find it. The signal needs resolution: coarse enough to know the archive exists, precise enough to return one item instead of everything.
Temple of the Dog was re-queried with new metadata.