Pocket Memory

Pocket Memory helps people save useful details and recall them with Meta Llama 3.2 3B running on Android. The revised narrated demo uses practical, fictional examples and actual app output: 1. Travel recall: ask where the spare keys were kept. Llama identifies the blue pouch in the front pocket of the grey backpack and cites the saved memory. 2. An updated client brief: Friday/English captions changes to Monday evening/Hindi captions. When a generated answer cannot be verified, the app explicitly shows the corrected source; earlier wording remains in history. The demo labels this safety fallback honestly. 3. Travel research: one real MCP search supplies official Google Maps Help excerpts, then local Llama answers how to download an area for offline use with source URLs. Private memories and earlier conversation are not attached to the web question. 4. Browser workflow: an actual WebMCP tool call returns Cedar Studio's current saved brief from the Android app. Setup stays inside Models: download Llama, verify the pinned file's bytes and SHA-256, then tap Use this model. The MCP pipeline allows one search and one local answer with bounded traffic, context, tokens, repetitions, cancellation and time. The browser companion uses phone loopback, a one-use code, expiring bearer session, same-origin checks and operation limits. A permitted browser agent may process returned content under its provider's policy. Pocket Memory's existing Android app and the public llama.cpp runtime are pre-existing foundations. New work is the Llama-only setup and model lifecycle, guarded MCP web-answer pipeline and WebMCP companion. Actual on-device inference, download/resume/Use, restart, live MCP, WebMCP and guard checks passed in an isolated Android emulator. Physical-phone testing and broader quality evaluation remain pending. The 2:05 demo is edited from actual captures with synthesized narration; it makes no speed claim.