Modes Explained

Modes Explained

SLM runs in one of three modes, and the mode answers one question: which language model, if any, may touch your memories. In Mode A, memories are saved and recalled without a language model. Mode B uses a model on your own machine. Mode C uses an endpoint you choose, which can be a cloud provider. In every mode the memory database stays on your computer, and you can change modes at any time without losing anything.

What each mode runs

The definitions live in one place, core/modes.py, so the dashboard, the CLI and the MCP tools all describe the modes with the same words. The one-line descriptions the server sends, and the dashboard shows, are these:

• Mode A: "Memories are saved and recalled without a language model — nothing leaves this device. Fastest and most private."
• Mode B: "A model on this machine — Ollama by default, any local server works — and nothing leaves this device."
• Mode C: "Your own endpoint, or a cloud provider, for the best recall quality. Your queries leave this device."

When the optional online answer check is switched on, the MCP description of Mode A and Mode B says "nothing else leaves this device" instead, because the check sends the question and top memories to its provider. The wording about Mode A is deliberately about saving and recalling: the on-device answer check (Answer Check) is a small separate model that runs after recall, on your own computer, and it does not change what Mode A means for your memories.

	Mode A: Local Guardian	Mode B: Smart Local	Mode C: Full Power
Language model	None. Extraction uses deterministic rules.	A model on this machine: Ollama by default, any OpenAI-compatible local server works.	Your own endpoint or a cloud provider.
What the model does	Nothing.	Fact extraction, entity resolution, type classification, importance scoring, answer generation.	The same as B, plus agentic multi-round retrieval and an optional cloud reranker.
Embeddings (preset)	Local nomic-ai/nomic-embed-text-v1.5, 768 dimensions.	Ollama nomic-embed-text, 768 dimensions.	Local nomic, 768 dimensions. Cloud embeddings only if you configure an endpoint (the catalogue default is text-embedding-3-small, 1536 dimensions).
Reranker	Local cross-encoder cross-encoder/ms-marco-MiniLM-L-12-v2.	Same.	Same, with semantic_top_k 80 and 2 agentic rounds.
LLM default	None.	ollama at http://localhost:11434, model gemma3:4b.	openrouter with openai/gpt-6-luna.

The presets are built in SLMConfig._mode_template (core/config.py). A fresh install writes all three mode files from these presets. After that, your own edits to a mode are kept. Mode C does not move you to a paid embedder on its own: a cloud embedding model is something you opt into, because a half-configured cloud embedder used to make ingestion fail.

What leaves the machine

This is the table people usually want, so it is deliberately blunt.

	Mode A	Mode B	Mode C
Memory text sent to a language model	Never.	Only to the model server you configured. If that is Ollama on another host, it goes there.	Yes. Extraction and consolidation text, and recall queries and context, go to the configured provider.
data_stays_local in the code	True	True	False

In Mode C the text that goes to the provider is the memory text in extraction prompts at save time, the query and top candidates during agentic recall refinement, and answer-generation prompts. Credentials in it are replaced first, but personal data is not.

Any mode can still send data off the machine when you turn on an optional feature: a custom embedding endpoint, a remote reranker, the online Answer Check, Web Access and Connected Apps, or encrypted Cloud Backup. SLM treats Mode A as "no language model in the save and recall path", not as "no network feature can ever be enabled". The code leaves the EU AI Act status of all three modes as None on purpose, because that needs a deployment assessment. See Privacy and Data Flow for the full picture.

Switching modes

There are four ways to switch, and they all end in the same function, SLMConfig.switch_mode.

slm mode          # show the current mode
slm mode b        # switch
slm restart       # apply it

You can also use the Settings pane, group "Operating Mode & LLM Provider", or the MCP tool set_mode (with get_mode to read it). To choose the model for B or C, use slm provider set <openai|anthropic|azure|ollama|openrouter|custom>. For your own OpenAI-compatible server, slm provider set custom --endpoint URL [--key KEY] [--model NAME] --mode b|c; the default target mode is C, and a keyless local server is fine.

A switch changes the mode field and nothing else. It saves your current settings to the old mode's file, loads the target mode's file if one exists, and otherwise copies your current settings and changes only the mode. LLM defaults are filled in only when you have no provider at all. Your embedding, retrieval, forgetting and math settings are not reset, and if a mode file names a different embedding model, set_mode answers needs_reindex and the daemon re-indexes in the background rather than blocking.

The reason for this design is an older bug where switching modes quietly replaced a tuned embedding endpoint with the mode's defaults. Switching now preserves what you set.

For an endpoint that is not on localhost, plain HTTP is accepted only for a numeric private-LAN address while retrieval.trust_plain_http_lan is true (the default). Public hosts and bare hostnames need HTTPS. Set the flag to false in a hardened network to require HTTPS everywhere off the machine.

Where a mode is stored

File in the data folder	What it holds
config.json	The active configuration, including mode.
mode_a.json, mode_b.json, mode_c.json	A saved copy of the settings for each mode.
current_mode	One letter, the single source of truth for the active mode.
answer_check.json, memory_kinds.json	Deliberately outside the per-mode copies.

Answer Check and memory-kind settings live in their own files because the per-mode copies are restored on every switch. Consent given in one mode would otherwise come back after you switched away and back. If config.json disagrees with current_mode, SLM loads the mode file that current_mode names.

What the default mode is

A fresh install starts in Mode A. SLMConfig.load() returns Mode A when there is no config.json, first-boot migration writes a into current_mode, and slm init --auto selects Mode A.

There is one edge case worth knowing. SLMConfig.read_current_mode() returns "b" when the current_mode file is missing or empty. Normally the file exists, so you will not see this, but if you delete it by hand while config.json still exists you can end up with two different answers to "what is the default". Recreate the file with slm mode a (or the letter you want) and the question goes away.

Embedding choices

The Settings pane offers two embedding providers: the default local model, or a custom OpenAI-compatible endpoint where you give a model name, dimension, endpoint and an optional key. From a terminal, slm embedder switch MODEL [--dimension N] [--provider sentence-transformers|ollama|openai|slm-media] [--endpoint URL] does the same, and keys are never taken on the command line.

slm models lists the catalogue (version 2026-10-08 in core/model_catalog.py): local and hosted language models and the embedders above, with test results for how often a model changed a fact during SLM's extraction test. The catalogue is advice. Any installed Ollama model, OpenRouter id or custom endpoint can be typed in.

Changing the embedding model re-indexes in the background

Vectors from different models cannot be compared, so a model change has to re-embed every memory. SLM does this inside the daemon without stopping: recall and remember keep using the old model and old vectors until the new space is complete, then both change in one transaction. The previous space is kept so one slm embedder rollback can return to it, and a daemon restart resumes the job where it stopped. A save that changes the vector width without going through this job is refused with HTTP 409 and the message points you to slm embedder switch MODEL --dimension N. The tuning variables are SLM_REINDEX_BATCH (32), SLM_REINDEX_PAUSE_S (0.02) and SLM_REINDEX_DUTY_REST (1.0). The full walkthrough, including why text stays on its own model, is on Embedding Models.

Language

There is no language or locale setting in SLM. Nothing in Settings, the setup wizard or SLMConfig selects a language, and the dashboard is English only (<html lang="en">). No catalogue entry is labelled multilingual.

What you can do for non-English text is replace two components. The bundled reranker, cross-encoder/ms-marco-MiniLM-L-12-v2, is English-only, and the code says so (core/config.py, retrieval/remote_reranker.py). Point SLM at a multilingual reranker such as BAAI/bge-reranker-v2-m3 or a Qwen reranker served by an OpenAI-compatible /v1/rerank endpoint, by setting retrieval.cross_encoder_backend to "openai" together with cross_encoder_endpoint and cross_encoder_model in config.json. This is not exposed in Settings. For embeddings, use a custom endpoint (for example Qwen3-Embedding) through Settings or slm embedder switch. Details and a worked JSON example are in Configuration.

SLM makes no claim about how well the default embedding model handles non-English text, because nothing in the code states it. Test it on your own material before relying on it.

Code: core/modes.py, core/config.py (SLMConfig, _mode_template, switch_mode, read_current_mode), core/model_catalog.py, core/embedding_reindex.py, cli/commands.py (cmd_mode), cli/embedder_cmd.py, mcp/tools_v3.py (set_mode, get_mode), retrieval/remote_reranker_config.py, core/provider_endpoint_trust.py.