Research Note: Offline Knowledge Is a Storage Budget
Published July 2026
Question
An offline AI terminal cannot "know everything" in the cloud sense. It can carry model weights, indexes, documents, and cached tools. Pockot needs to measure how much useful reference material fits before storage, update, and retrieval limits dominate the device.
Source-Backed Data Points
- Wikimedia's Kiwix ZIM index lists `wikipedia_en_all_maxi_2026-02.zim` at 123,980,647,016 bytes and `wikipedia_en_all_nopic_2026-06.zim` at 52,690,706,555 bytes. Source: Wikimedia Kiwix ZIM index.
- The same index lists `wikipedia_en_medicine_maxi_2026-04.zim` at 2,215,520,569 bytes and `wikipedia_en_medicine_nopic_2026-04.zim` at 862,376,216 bytes. Source: Wikimedia Kiwix ZIM index.
- Meta describes Llama 3.2 1B and 3B as models built for edge and mobile devices. Source: Meta Llama 3.2.
Reading
Offline knowledge is not only a model parameter count. A small model may fit in memory, but the device still needs reference files, vector indexes, metadata, update logs, and source boundaries. A 256 GB device can carry a large offline encyclopedia snapshot, but it cannot carry every raw corpus, every embedding, and every media file without tradeoffs.
The useful design split is model memory versus knowledge storage. Model weights affect runtime and RAM. Corpus files affect disk and retrieval latency. Indexes affect both storage and quality. Updates affect trust because stale knowledge can be worse than missing knowledge when the device is used as a reference tool.
Pockot should therefore avoid the claim that a pocket device "contains the internet." It can contain curated source packs, offline encyclopedias, manuals, and local notes. The stronger product question is how to choose the reference set and show its date, size, license, and retrieval method.
Tool Rule
The feasibility calculator should keep `corpus_gb` visible and add named corpus profiles. A model that fits but leaves no space for a reference set should not be labeled as a useful offline assistant configuration.