Before AI model calls
Local safety gate
Bella checks each chat message locally before memory retrieval, embeddings, tools, or OpenAI/provider calls. Unsafe prompts are stopped before they leave Bella's backend.
- Blocks prompt-injection attempts, hidden-instruction requests, and secret extraction.
- Blocks malware, credential theft, fake-document, fraud-evasion, weapon-harm, and oversized requests.
- Returns a safe refusal without calling the model provider.
Read the AI safety articleBy subscription plan
Memory and context controls
Bella limits how much memory and recent context each plan can inject into a turn. She also checks stored-memory limits before creating embeddings or saving new long-term memory.
- Starter gets lighter context, higher plans get deeper recall.
- Chat continues when memory is full, but extra saves can be skipped.
- Near-limit and full-limit states can show upgrade or add-on prompts.
Read the memory limits article