Settings

Manage model access tokens and research workstation preferences.

Hugging Face Access Token
Required to download gated embedding and language models (like litert-community/embeddinggemma-300m) directly into your browser cache.
Need a token?Generate one on Hugging Face(Only "Read" permissions needed)
Your token never leaves this device. It is stored securely in your browser's IndexedDB database and sent exclusively to Hugging Face CDN endpoints during model download.
LLM Humanizer
Disabled
Steers on-device Gemma 2B decoding and prompts for natural phrasing, varied cadence, and cliché elimination.
Extended Context Window (RLM)
WASM Engine
Bypasses prompt truncation limits (arXiv:2512.24601) by ingesting complete 50KB–200KB+ academic documents into in-browser WebAssembly linear memory and orchestrating recursive sub-calls.

When enabled, synthesis and deep paper questions execute iterative REPL turns directly inside your browser sandbox rather than cutting text to a 16KB window. Unlocks context configurations up to 128K tokens.

Zero remote servers. The WebAssembly sandbox operates entirely client-side using @buddhilive/sandbox and local LiteRT WebGPU/WASM model weights.
Context Window & Memory Optimization
RLM Scalable
Configure the maximum token capacity for conversation history. When conversations reach 80% of this threshold, older turns are automatically compacted into high-fidelity memory to avoid KV-cache overflow.
Context Window Limit

Scalable up to 128K tokens powered by in-WASM RLM sandbox.

4k tokens

(4,096 tokens)

1k tokensThreshold: ~3.2k tokens (80%)128k tokens
Automatic Chat Compaction

When conversation tokens exceed 3,276 tokens (80% capacity), the system non-destructively summarizes intermediate conversation turns. Your full transcript remains visible in chat history and IndexedDB, while the active LiteRT-LM inference prompt buffer stays lean and fast.

RLM mode active: Long context summaries are generated hierarchically inside the WebAssembly sandbox.