Full history makes the agent dumber. apple's shared selective persistent memory research, posted…
full history makes the agent dumber. apple's shared selective persistent memory research, posted september 16, found naive full-history persistence actively degrades task completion by biasing agents with stale reasoning traces, while selective memory succeeded in 12 of 12 zero-token refreshes.
the brain needs a delete key.
Context
The paper Shared Selective Persistent Memory for Agentic LLM Systems (arXiv 2607.09493, authors with Apple affiliations) compares no memory, full history and shared selective memory on 24 recurring artifact-generation tasks with blinded raters. Mean completion was 79 percent with no memory, 71 percent with full history and 96 percent with selective memory; input tokens 2.1K, 18.7K and 3.4K. It reports 100 percent completion with zero LLM tokens across all 12 selective-memory refresh trials, on schema-compatible public data. The authors motivate it by saying a prior session's trace is potentially misleading.
It is a small preprint by the authors' own protocol, with per-use-case sample sizes such as 10 and 8. Stale reasoning traces as the cause is the paper's explanation and not an isolated causal test in the text read. The 12 of 12 result applies to a narrow condition of compatible public data. The arXiv identifier and a July 10 aggregator date point to July 2026, and September 16 was not supported. The brain needs a delete key is the author's take.
Related work
- Earlier note on the same Apple paper ↗Same paper and claims.
Watch next
- A larger or public-benchmark replication.
Sources
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 21 September 2026 at 14:24 IST. Sources are the papers and datasets the note draws on.
View the original post ↗Embed this note
More notes
The air is now being asked to keep its own ledger
the air is now being asked to keep its own ledger: ecmwf’s aifs compo becomes the first ai model to forecast atmospheric composition globally every three hours, cleanair simulates 365 days of pm2.5 over china in ten seconds, and a unified framework maps six pollutants at one kilometer across the whole country. the air now files its own composition report.
read the note →The current is now being asked to draw its own map
the current is now being asked to draw its own map: china’s langya 2.0 predicts six ocean phenomena including internal waves and mesoscale eddies, a deep net called wenhai resolves eddies globally with air sea flux formulas built in, and scripps infers surface currents from the way temperature patterns deform in satellite images. the ocean now files its own circulation report.
read the note →The soil is now being asked to report its own carbon
the soil is now being asked to report its own carbon: a nix color sensor paired with generative data augmentation predicts soil organic carbon without a lab, random forest drives 74 percent of soil health mapping studies, and sentinel 2 tracks five year carbon change across france and italy from 922 samples. the dirt now files its own carbon account.
read the note →