Optional persistent wiki memory

Off by default. A background model maintains factual entries, summaries, evidence quotes and links. New turns retrieve relevant entries using triggers, BM25 and up to two graph hops. No additional embedding model or vector database.

Real model recall with no recent chat

The test taught Bob/Alice, corrected Bob’s home from Hamburg to Bremen, and used a fresh session with a zero-token recent-chat budget. All examples below are authored synthetic test data.

QuestionReplyRetrievalPass
Wie heißt Bobs Freundin?Bobs Freundin heißt Alice.48.65 msTrue
Wo wohnt Bob jetzt?Laut meinen Daten wohnt Bob in Bremen.0.94 msTrue
Was macht Alice gern?Alice mag Jazz und den Garten.0.9 msTrue

Background update times: 732.38 ms, 525.78 ms, 595.5 ms, 798.04 ms. These additional provider calls run after the reply; they do not delay first audio. Measurements are a small functional test, not a universal retrieval-quality benchmark.

Browser and isolation checks

{
  "passed": true,
  "default_off": true,
  "recall_without_recent_history": true,
  "disable_verified": true,
  "reset_verified": true,
  "off_persists": true,
  "mobile_fits": true,
  "errors": []
}

Limits and controls

Default ceilings: 10,000 knowledge tokens and 5,000 recent-chat tokens, using conservative UTF-8 accounting. The local model has a smaller combined allowance. Memories are untrusted reference data; a literal quote establishes provenance, not truth. Extraction and ASR mistakes remain possible. Inspect each entry’s evidence and five saved revisions. Separate browser capabilities isolate workspaces; this is not biometric authentication or account-based authorization.

The first test exposed a lexical retrieval miss for “Bobs” versus “Bob”; possessive-name expansion fixed it. Multiple interests under one field are preserved. The failed first run remains in the earlier receipt.

Reset deletes wiki entries, versions and saved/current model chat context. An epoch guard prevents an older pending update from recreating deleted facts. Audio speaker profiles and camera settings remain separate. Server files are private filesystem text, not encryption at rest.

Inspect and reproduce

Full architecture, privacy, setup and commands · Model receipt · Browser receipt · Quick-reaction latency study