> ## Documentation Index
> Fetch the complete documentation index at: https://docs.memloom.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Configuration

> config.env, the two modes, and the embedding fingerprint

## Where config lives

All configuration is one dotenv-style file, created by `memloom init`:

```
~/.memloom/config.env
```

The daemon reads it **at startup**. After editing it, restart (`memloom stop`, then any
command). Real environment variables take precedence over the file, so a shell-provided or
MCP-provided key overrides what's written there. Set `MEMLOOM_HOME` to relocate the whole
home directory (config *and* data).

## Reference

```bash theme={null}
# The one switch that matters: with a key, memloom runs in cloud mode;
# without it, offline mode.
OPENROUTER_API_KEY=sk-or-...

# Optional overrides (defaults shown):
OPENROUTER_EMBED_MODEL=qwen/qwen3-embedding-8b
OPENROUTER_EMBED_DIMS=1024          # Matryoshka truncation from the native 4096
OPENROUTER_EMBED_PROVIDER=nebius    # preferred OpenRouter host for embeddings
OPENROUTER_LLM_MODEL=google/gemini-2.5-flash
OPENROUTER_CHAT_MODEL=google/gemini-2.5-flash  # assistant chat; must support tool calling
MEMLOOM_AUTO_INDEX=on               # background entity extraction after saves/ingests
MEMLOOM_INDEX_CONCURRENCY=6         # entity-extraction worker pool size

# Notion connector (see the Notion page in the CLI reference):
NOTION_TOKEN=ntn_...                # internal integration token; pages shared with it sync
NOTION_POLL_MS=300000               # how often the daemon checks for edits (min 60000)

# Storage tier: unset = embedded PGLite in ~/.memloom/data. Set to run on a real
# Postgres server with pgvector instead (see the self-hosting guide):
MEMLOOM_PG_URL=postgres://user:password@localhost:5432/memloom
```

<Note>
  `MEMLOOM_AUTO_INDEX` controls the background indexing that runs a few seconds after a
  save or file ingest (one LLM call per new item, debounced into batched runs). Set it to
  `off` if you want every LLM call to be explicit; `memloom index` and the Console still
  work either way. Offline mode never auto-indexes.
</Note>

<Note>
  `OPENROUTER_EMBED_PROVIDER` exists because embedding latency varies wildly between hosts of
  the same model on OpenRouter: the same call has been observed taking 16s on one host and
  under a second on another. For the default model, memloom pins the fast host (nebius)
  automatically; set this only if you change models or want a different host.
</Note>

## The two modes

|                                   | **Cloud** (key set)           | **Offline** (no key)                               |
| --------------------------------- | ----------------------------- | -------------------------------------------------- |
| Embeddings                        | Real model via OpenRouter     | Deterministic hash embeddings                      |
| Recall quality                    | Full semantic                 | Word-overlap only (keyword arm still full quality) |
| Dedup / versioning on save        | ✓                             | ✗ (every save is `added`)                          |
| Conflict detection                | ✓                             | ✗                                                  |
| `memloom index` (entities, graph) | ✓                             | ✗ (errors; needs the LLM)                          |
| Context ingestion & citations     | ✓                             | ✓                                                  |
| Network calls                     | Embedding + LLM requests only | None                                               |

Offline mode is for trying memloom, tests, and air-gapped storage of text you'll search by
keywords. For real use, set the key.

## The embedding fingerprint

A store's vectors are only comparable to vectors from the same provider + model + dims. On
first init, memloom stamps the store with the embedding fingerprint (e.g.
`openrouter:qwen/qwen3-embedding-8b@1024`). Any later startup with a **different**
fingerprint is refused with an explicit error, because mixed embedding spaces don't fail
with an error; they silently return garbage recall.

Switching between offline and cloud mode (or changing embedding models) is a migration, not
a dead end: stop the daemon and run [`memloom reembed`](/cli/graph#reembed), which recomputes
every vector with the new provider and stamps the new fingerprint. The LLM model, by
contrast, can be changed freely; it doesn't touch the vector space.

## Daemon endpoints

| What              | Where                                            | Change                                                                                                                                                                     |
| ----------------- | ------------------------------------------------ | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| HTTP API + viewer | `http://127.0.0.1:4319`                          | fixed (localhost-only by design)                                                                                                                                           |
| Postgres wire     | `postgresql://postgres@127.0.0.1:54329/postgres` | fixed; distinctive port so it never collides with a local Postgres on 5432. Embedded tier only: with `MEMLOOM_PG_URL` set there is no bridge, inspect your server directly |

Both bind to `127.0.0.1` only, and the API's CORS policy allows only localhost origins, so a
public web page can't drive your daemon.