Start with a sentence
Eligible user text updates private word-sequence counts when learning is enabled. Duplicate messages do not increase their weight. Assistant outputs are not fed back as training examples.
The silver comet glows brightly above the quiet valley.Inspect the next word
Predictions combine bundled wording, explicitly shared examples and private text. The same state and input give the same result. The percentages describe the next token, not factual certainty.
next word: the silver cometComplete a phrase
Completion follows short matching word sequences and stops when context is too weak or repeats. It may be awkward or incorrect; review its wording. It does not replace computed answers.
complete: the silver cometInspect word importance
A transparent heuristic scores repeated words against their rarity in the bundled corpus. Rare words can be useful search terms, but rarity does not prove meaning or truth.
word importance: The silver comet glows above the valley.Retrieve what you pasted
Pasted prose remains private and can supply quoted answers later, including in another session under the same identity. Ask about specific words or subjects. Excerpts are attributed to your earlier text.
Learn dialogue examples
User:/Assistant: and Q:/A: pairs can be pasted as source lessons. These role labels are data, not privileged instructions. Exact matching retrieves the supplied answer.
User: What is the workshop greeting?
Assistant: Welcome back. What shall we build today?Reason over premises
The English class-logic reader supports all/every inclusions, no exclusions and is a/is not a membership. It follows bounded chains, retains supporting premises and reports contradictions. It does not infer the converse or treat unknown as false.
All robots are machines. Ava is a robot.
Is Ava a machine?Inspect and control
Review learned counts and observed prediction accuracy. Accuracy is measured before each new message is trained and is not an independent benchmark. Pause learning disables automatic collection; clear language learning separately from transcripts.
language learning
learn from my past chats
forget language learningShare only public examples
The explicit share language command publishes only its supplied text for other users word predictions. Do not include private or confidential wording. Shared prose does not become verified knowledge or change tool permissions. Published examples are separate from private workspace deletion; the operator can remove them.
share language: Clear explanations connect a conclusion to the evidence that supports it.Understand capacity
Private storage accepts up to 200 distinct text samples, 250,000 characters and 20,000 word transitions. Oldest recall samples roll out as new text arrives; counts decay every 64 new samples. At most 2,500 tokens are processed per message. Imported samples rebuild counts; observed accuracy history is not restored. Source recall and proof searches have explicit limits. This is a small statistical language model, not neural training or AGI.
Read the methods and limitations or try the examples in the workspace.