Minimum connection settings
Mneme uses an OpenAI-compatible endpoint for LLM-backed work. Set these values in .env, or let the onboarding wizard save them for you.
API_KEY=sk-your-api-key-here\nBASE_URL=https://api.deepseek.com/v1\nLLM_MODEL=deepseek-chatEndpoint boundary
Remote endpoints must use HTTPS by default. Plain HTTP is permitted for loopback addresses. A non-local HTTP endpoint requires MNEME_ALLOW_INSECURE_HTTP=1.
Retrieval and generation
| Variable | Default | Purpose |
|---|---|---|
LLM_TEMPERATURE | 0.2 | Generation temperature. |
LLM_TOP_K_MIN | 12 | Minimum chunks for standard retrieval. |
LLM_TOP_K_MAX | 70 | Maximum chunks for standard retrieval. |
ALPHA | 0.7 | Graph RAG fusion weight. |
EMBEDDING_MODEL_NAME | all-MiniLM-L6-v2 | Model identifier for local loading. |
Resource limits
| Variable | Default | Purpose |
|---|---|---|
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 retrieved context sent to an endpoint. |
MNEME_DOCUMENT_ROOT | unset | Optional root allowed for indexed files. |
What can leave the machine
File discovery, parsing, indexing, cache management, and vector/BM25 retrieval run locally. When an LLM-backed feature is used, retrieved snippets are sent to your configured endpoint for answer generation, query decomposition, or Graph RAG entity extraction.