Mneme/Reference

Reference

Environment reference

Controls for endpoint access, retrieval, model loading, file roots, and bounded remote context.

Endpoint and model

Variable	Default	Description
`API_KEY`	—	OpenAI-compatible API key.
`BASE_URL`	`https://api.openai.com/v1`	LLM endpoint; remote endpoints require HTTPS.
`LLM_MODEL`	`deepseek-chat`	Chat and query-decomposition model.
`LLM_TEMPERATURE`	`0.2`	Generation temperature.
`EMBEDDING_MODEL_NAME`	`all-MiniLM-L6-v2`	Embedding model identifier.

Retrieval

Variable	Default	Description
`LLM_TOP_K_MIN`	`12`	Minimum retrieved chunks.
`LLM_TOP_K_MAX`	`70`	Maximum retrieved chunks.
`ALPHA`	`0.7`	Graph RAG fusion weight.
`RAG_WATCH_DIR`	—	Directory watched by the TUI.

Safeguards

Variable	Default	Description
`MNEME_DOCUMENT_ROOT`	—	Optional allowed root.
`MNEME_MAX_DOCUMENT_BYTES`	`52428800`	Maximum document size, 50 MiB.
`MNEME_MAX_PDF_PAGES`	`2000`	Maximum PDF pages.
`MNEME_MAX_REMOTE_CONTEXT_CHARS`	`60000`	Maximum remote context.
`MNEME_ALLOW_INSECURE_HTTP`	unset	Explicit non-local HTTP override.

Read with the product boundary

These controls do not make indexed documents trusted instructions. Mneme treats retrieved text as data and keeps source metadata in the answer context.