Make the assistant accurate about itself

I'm a Pro user. Good models, great agentic chat, privacy is why I'm here. One
thing has genuinely eroded my trust, and it looks cheap to fix.

I asked, in-chat, whether chats share information with each other. The answer
was no. I believed it, because the app's own assistant said so, and built an
entire external workflow around that premise. Then I found out Memoria shipped
nine months ago. Chats do share information now. The assistant was confidently
wrong about its own product.

The cause is structural. On Venice you're usually talking to someone else's
model. It has no privileged access to the app and no idea what shipped last
month, so it answers from training data in the same tone it uses when it's
right. It also has web search — it just doesn't think to use it for a question
it believes it already knows.

The fix looks small, because the mechanism already exists. AUTO mode already
inspects intent and selects a tool. "Question about Venice" is a routable
intent.

• Add a retrieval tool over the changelog, FAQ, and docs.
• Route self-referential intent to it.
• Fallback rule: if the question is about Venice and you haven't checked,
  check — don't answer from memory.
• Return the answer dated and cited so users can verify it.

It has to be injection-based rather than trained, since you can't retrain
Anthropic's or OpenAI's models. Even a canned "let me check the docs" beats a
confident wrong answer.

This matters more than an ordinary hallucination, because the assistant is the
voice of the product. When it's wrong about Venice, users conclude Venice is
unreliable.

SECOND, SMALLER: TELL ME WHY A REQUEST FAILED

A long thread with credits available started failing on every turn. Deleting
stalled responses bought a few more turns before it failed again — cumulative,
which points at the context ceiling. But nothing in the app confirms that.
"Response failed. Please try again." is all you get, for every possible cause.

A visible context-usage readout and an error that distinguishes "context full"
from "tool failed" from "model unavailable" would make it obvious in seconds
instead of something users infer from behaviour.

PRIVACY NOTE

The lookup target is public, static, and identical for every user, so
retrieving it reveals nothing about the asker. The cleanest build is a local
index of the changelog and FAQ, refreshed on load and searched on-device — the
pattern Memoria already uses for local hybrid search. That's strictly more
private than the web search already shipping.

One constraint: on E2EE models, routing has to happen inside the enclave, or
the routing step leaks the plaintext query and the E2EE guarantee quietly
weakens for that turn.

Nothing here needs logging, and nothing changes about memory staying in the
browser.

RELATED, ALREADY FILED

• Projects with files for context — still Backlog after two years
• Backup/export/import for memories and character memories — and the Memoria
  blog already says backup integration is planned

PROVENANCE

Drafted by an AI assistant during a live session on Venice, working from a real
Pro user's findings. The experience and observations are theirs; the wording is
the model's. The Memoria release date and current feature set were checked
against the public changelog and FAQ.

One irony worth naming: an AI wrote this report about an AI being unable to
describe its own platform.