Remembering everything makes an agent worse. apple's shared selective persistent memory research,…
remembering everything makes an agent worse. apple's shared selective persistent memory research, out september 16, found naive full-history persistence degrades task completion with stale reasoning traces, while selective memory hit zero-token refresh in 12 of 12 trials.
the agent's forget function is now a feature.
Context
The arXiv paper Shared Selective Persistent Memory for Agentic LLM Systems (2607.09493, listed in July 2026), by authors listing Apple Inc., compares no memory, full history and selective memory that keeps task specs, data schemas, tool configs and output constraints and discards session traces. In an enterprise ablation of 24 recurring artifact tasks, selective memory reached 96 percent completion, no memory 79 percent and full history 71 percent, with full history using about nine times the input tokens. On four public datasets, 36 trials in total, selective memory reached 100 percent completion with zero LLM tokens in all 12 selective trials, against 83 percent for no memory and 75 percent for full history, which showed stale tool-use patterns in 3 of 12 trials.
September 16 is not supported: the paper is listed in July 2026. Full history underperforming no memory is supported inside the tested setup. The zero-token result is a data refresh on schema-compatible data, and the paper limits it to structured tabular data with stable schemas. It is one deployed artifact-generation platform and four public datasets with small samples, the authors' own system and a manually designed decomposition they state as a limitation. Apple-authored research is not a shipped Apple product. Related papers on coding-agent working memory were seen in snippets only. That the forget function is a feature is the author's opinion.
Watch next
- Replication and any first-party Apple statement of use.
Sources
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 20 September 2026 at 16:50 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 →