Configuration

Configuration

SLM keeps all of its settings in plain files inside one data folder, and almost every setting has a command, a Settings group in the dashboard, or both, so you rarely edit JSON by hand. This page explains where the folder is, which file holds what, which Settings group changes which behaviour, and every environment variable that was confirmed in the source, with its default and where it is read.

The data folder

The data folder is ~/.superlocalmemory/ unless you move it. The environment variables below are checked in this order and the first one that is set wins, followed by a base directory passed in explicitly, then a base_dir hint in an old config.json, then the default. The resolver only reads; it never moves data.

1. SLM_DATA_DIR
2. SL_MEMORY_PATH (older name)
3. SLM_HOME (older name)

Identity files (.install_token, .signer_key, api_key, credentials.json) are treated as durable across a move of the folder. If the folder you select and the default folder both already hold data, startup is blocked with a conflict error, and SLM will not merge or move either one for you. The reason is that two daemons serving two databases is how stores get damaged, so SLM makes you choose.

File or folder	What it holds
memory.db	Memories, facts, entities, graph, scenes, mesh, access control, audit, evolution log, corrections.
learning.db	Learning signals, ranking state, receipts, Answer Check history, saved views.
media.db, media/	Pictures, documents, folder sources and the original files.
config.json	The active SLMConfig: mode, profile, LLM, embedding, retrieval, scope, forgetting, evolution and the rest.
mode_a.json, mode_b.json, mode_c.json, current_mode	A saved copy of the settings for each mode and the letter of the active one. See Modes Explained.
profiles.json	The profile list and active profile, kept in step with the database. See Profiles.
answer_check.json, memory_kinds.json	Settings kept outside the per-mode copies so a consent cannot come back after a mode switch.
optimize.json	Optimize settings, hot-reloaded every 2 seconds.
code_graph_config.json	Code-graph switches.
features.json	Whether images, documents and folder sources are turned on.
adapters.json	Ingestion adapter configuration. See Ingestion Adapters.
backup_config.json, cloud-backup-encryption.json, backups/	Backup schedule, encryption state and the local copies. See Cloud Backup.
remote_keys.json	Web Access keys (stored as digests).
store-check.json	Result of the memory-store check.
config.toml	Optional installer-written file. SLM reads only its [deployment] section.
.disabled	Global off switch, created by slm disable.

How the layers combine

Settings reach config.json from the Settings pane, from commands such as slm mode, slm provider, slm embedder and slm config set, or from you editing the file. config.json is written atomically and forced to owner-only permissions (0600) because it can hold a provider key. The environment variables in the tables below are read where they are used and are never written back into config.json, so a process-only override cannot leak into the file on a later save. Prefer the environment for secrets such as SLM_CROSS_ENCODER_API_KEY or the provider keys, because that avoids a plaintext copy on disk.

If config.json is unreadable or corrupt, SLM falls back to the Mode A default and logs a warning rather than refusing every command.

slm config get <dot.key> reads any value. slm config set <dot.key> <value> is deliberately limited to this list: evolution.enabled, evolution.backend, evolution.max_evolutions_per_cycle, evolution.mutation_model, evolution.verify_model, evolution.confirm_model, mesh_enabled, daemon_idle_timeout, entity_compilation_enabled, graph_backend, vector_backend, scope.default_scope, scope.recall_include_global, scope.recall_include_shared. scale_engine_state is excluded on purpose: it records what was done to the store, so writing it by hand could make the daemon skip or repeat a projection.

The Settings pane

The Settings pane has a search box over every setting and 16 groups. "Save all changes" at the bottom clicks the Save button of the groups that have one, and each group says whether a daemon restart is needed.

Group	What it changes
Operating Mode & LLM Provider	Mode A, B or C; provider (None, Ollama, OpenRouter, OpenAI, Anthropic), model, API key, base URL; a Test button.
Answer check	On this Mac (Laya), Online with Jev, or Off; setup, repair, test, remove. See Answer Check.
Embeddings & Retrieval	Default local model or a custom OpenAI-compatible endpoint (model, dimension, endpoint, key, Test); the background re-index panel; and, only for a store whose text an older release moved to the picture model, the memory-engine card with Roll back. See Embedding Models.
Storage Backends	Graph backend auto, sqlite or cozo; vector backend auto, sqlite-vec or lancedb; base directory (read-only). Saving reports whether a restart is needed.
Memory Scope (auto-saves)	Default write scope (personal, shared, global) and whether recall includes shared and global memories. Both recall switches are off by default.
Recall & Memory Behaviour (auto-saves, no restart)	Recall depth (retrieval.top_k, 1 to 200, shown as 20), the reranker switch (about 200 MB of RAM and about 200 ms), memory injection into prompts, the core-memory block and its size (0 to 50, default 5).
Auto-Capture	Capture from agent sessions; capture decisions; capture bug fixes.
Auto-Recall	Inject relevant memory into agent context; recall on session start.
Auto-Invoke (FOK Threshold)	Run recall automatically above a relevance threshold (0 always, 1 never); ACT-R mode.
Forgetting Engine	Enable; protect core memories; archive and forget thresholds; run interval in minutes; the trust-based fade factor (kappa) with a decay preview.
Skill Evolution	Off by default. Enable, and choose a backend (auto, claude, ollama, anthropic, openai). It makes background model calls. See Skill Evolution.
Backup	Auto-backup on or off, daily or weekly, copies kept, Backup Now, link to Cloud Backup.
Mesh Network	mesh.enabled.
Trust & Source Quality	Weight recall by trust; prefer first-party agents; minimum trust to promote a memory.
Rate Limits	Write limit, read limit and window in seconds. Applied immediately; the local dashboard gets a multiple of the limits.
Daemon (port changes require restart)	API port, legacy port, whether the legacy port is open, idle timeout (0 = run forever).

Not in Settings: the remote reranker (config.json only), retention policy and company mode (Governance pane), memory-kind settings (slm kinds settings), Optimize (its own pane), Web Access (slm remote), the code graph, and turning on images and documents (the Documents & Images pane). The limits you set under Rate Limits are saved to config.json and re-applied at daemon start.

Environment variables

Every variable below was found by searching the source. The default is stated only where the code states it. "Source" is where the variable is read.

Where SLM keeps things and who is allowed to write

Variable	Effect	Default	Source
SLM_DATA_DIR, SL_MEMORY_PATH, SLM_HOME	Data folder, first one set wins.	~/.superlocalmemory	infra/data_root.py
SLM_AGENT_ID	Agent name attached to MCP calls on stdio.	mcp_client	mcp/agent_context.py
SLM_USER_SESSION	A dashboard user session the CLI sends to the daemon in X-SLM-User-Session, for governed workspaces.	unset	cli/daemon.py
SLM_PII_REDACTION	A truthy value turns on PII redaction when saving (also the pii_redaction config key and the deployment policy).	off	memory_core/save_path.py
SLM_REQUIRE_CREDENTIALS	1 makes even loopback writes present a daemon capability, install token or API key.	off	server/write_identity.py
SLM_REQUIRE_API_KEY_LOOPBACK	When set and an api_key file exists, loopback writes must also send a matching X-SLM-API-Key.	off	infra/auth_middleware.py
SLM_DISABLE	Any value other than 0, false, no or off makes SLM do nothing, like the .disabled file.	unset	core/slm_disabled.py
SLM_NON_INTERACTIVE	Makes the setup wizard treat the session as non-interactive.	unset	cli/setup_wizard.py
SLM_PYTHON	Python interpreter the npm installer uses; it must be 3.12 to 3.14.	searched automatically	scripts/postinstall.js

MCP

Variable	Effect	Default	Source
SLM_MCP_PROFILE	Tool profile: core, code, full, power, mesh, whole. See Profiles.	the default set, equal to full	mcp/server.py
SLM_MCP_TOOLS	Comma-separated allowlist of tool names.	unset	mcp/server.py
SLM_MCP_ALL_TOOLS	1 registers every tool.	unset	mcp/server.py
SLM_MCP_MESH_TOOLS	1 or true exposes the mesh tools even if mesh is off in config.	follows mesh_enabled (on)	mcp/server.py
SLM_MCP_STATEFUL	1 opts out of stateless Streamable HTTP and uses session ids.	stateless	core/remote_mode.py
SLM_DISABLE_WARMUP_SIDE_EFFECTS	1 skips the dependency check and warm-up side effects at import.	unset	__init__.py

Daemon, memory and request limits

Variable	Effect	Default	Source
SLM_DAEMON_PORT	Daemon port.	8765	server/unified_daemon.py
SLM_DAEMON_HOST (or SLM_HOST)	Address the daemon binds. Anything beyond loopback logs a warning unless SLM_REQUIRE_CREDENTIALS=1.	127.0.0.1	server/unified_daemon.py
SLM_DAEMON_IDLE_TIMEOUT	Seconds idle before the daemon stops; 0 never. Falls back to daemon_idle_timeout.	0	server/unified_daemon.py
SLM_DISABLE_LEGACY_PORT	1 or true turns off the 8767 redirect.	on	server/legacy_port.py
SLM_RSS_BUDGET_MB	Total memory budget for the health monitor; 0 computes 40 percent of RAM.	0	server/unified_daemon.py
SLM_MAX_WORKER_MB	Per-worker size limit checked by the memory watchdog.	2500	server/unified_daemon.py
SLM_EMBED_WORKER_RSS_LIMIT_MB	Resident-size cap for the embedding worker.	2500	core/embedding_worker.py
SLM_RATE_LIMIT_WRITE, SLM_RATE_LIMIT_READ, SLM_RATE_LIMIT_WINDOW	Per-client request limits and window in seconds. Loopback gets max(300, 10 x write) writes and max(2000, 20 x read) reads.	30, 120, 60	core/remote_mode.py
SLM_DB_BUSY_TIMEOUT_MS	How long a database call waits for a writer.	10000	storage/database.py
SLM_LOOP_KILL	Any non-empty value stops bounded loops.	unset	loops/engine.py

Web Access and remote use

Variable	Effect	Default	Source
SLM_REMOTE	Opts the daemon into LAN and distributed mode.	off	core/remote_mode.py
SLM_MCP_ALLOWED_HOSTS	Comma-separated trusted clients: *, IP, CIDR or prefix wildcard.	unset	core/remote_mode.py
SLM_ALLOWED_HOSTS	Host names the daemon accepts in the Host header, as a DNS-rebinding guard; *.suffix allowed.	localhost and IP literals	server/host_guard.py
SLM_REMOTE_LISTEN	Address for the TLS remote listener; overrides the listen value in remote/remote.json.	unset	server/remote_listener.py
SLM_REMOTE_TLS_CERT, SLM_REMOTE_TLS_KEY	Certificate and key for that listener.	unset	server/remote_listener.py
SLM_REMOTE_ALLOW_PLAINTEXT	1 allows remote callers over plain HTTP; SLM logs that keys and memory are then readable on the network.	off	server/remote_access.py

Retrieval, ranking and caching

Variable	Effect	Default	Source
SLM_CROSS_ENCODER_API_KEY	Bearer token for a remote rerank endpoint; overrides the config value.	unset	retrieval/remote_reranker.py
SLM_REMOTE_RERANK_DEADLINE_S	Time limit for one remote rerank call, retries included.	1.0	retrieval/remote_reranker.py
SLM_RERANKER_ENABLED	Process-only override of use_cross_encoder; 0, false, no, off disable.	the config value	core/config.py
SLM_RANKING	Adaptive ranking: v1, v2 or v2-ensemble.	off	core/recall_pipeline.py
SLM_RECALL_NO_SUPPORT_GATE	1, true or yes turns the support check off and restores the older recall behaviour; the setting retrieval.answer_support_gate set to false does the same.	on	retrieval/support_gate.py (switched_off)
SLM_SKIP_LAYA	1, true, yes or on stops setup and upgrades installing the on-device answer check; slm setup --no-laya does the same for one run.	off	cli/laya_offer.py (opted_out)
SLM_BANDIT_DISABLED	1 skips bandit re-ranking and its background loops.	off	core/recall_pipeline.py
SLM_TRUST_READ_GATE	1 enables the trust check on reads.	off	trust/gate.py
SLM_CACHE_BACKEND	tiered or sqlite for the derived cache.	tiered	cache/factory.py
SLM_CACHE_DB	Path of the active-brain cache database.	active_brain_cache.db in the data folder	core/context_cache.py
SLM_DERIVE_CACHE_MAX_MB	Size cap of the derived cache.	2048	cache/sqlite_store.py
SLM_AUTO_PROMOTE_MIN_EDGES	Graph edges before the scale engine auto-promotes.	100000	core/backend_orchestrator.py
SLM_VECTOR_HISTORY_DAYS	How many days of old vector-store versions the background compaction keeps. Every saved memory writes a version, and every semantic search walks the versions, so a long history slows search.	1/24 (one hour)	core/maintenance_scheduler.py
SLM_REINDEX_BATCH, SLM_REINDEX_PAUSE_S, SLM_REINDEX_DUTY_REST	Batch size, pause between batches in seconds, and rest as a fraction of batch time during a model switch.	32, 0.02, 1.0	core/embedding_reindex_steps.py
SLM_EVOLUTION_BACKEND	claude-cli runs skill evolution through the Claude CLI.	Anthropic API	evolution/llm_dispatch.py
SLM_EVO_AUTO_APPROVE	1 or true auto-approves skill evolutions. EvolutionConfig has no field for it, so the variable decides.	off	evolution/skill_evolver.py

SLM_EVOLUTION_ENABLED also exists, but the daemon sets it for its hooks from evolution.enabled, so set the config key instead.

Pictures and documents

Variable	Effect	Default	Source
SLM_MEDIA_ALLOW_LOW_RAM	1 lets images and documents turn on below the 15 GiB check.	off	runtimes/media_env.py
SLM_MEDIA_WORKER_RSS_LIMIT_MB	Overrides the picture worker's size cap.	per model (4500 MB for EmbeddingGemma 2)	runtimes/media_models.py
SLM_MEDIA_SPACE_MODE	paired or separate picture search space.	separate	runtimes/space_plan.py
SLM_DOC_MAX_MB	Largest PDF accepted, in MB.	100	documents/submit.py
SLM_DOC_MAX_PAGES	Most pages in one PDF.	500	documents/runner.py
SLM_MEDIA_URL_HOSTS	Extra hosts a connected web app may fetch attached files from.	none	core/media_fetch.py

Mesh

Variable	Effect	Default	Source
SLM_MESH_HOST	Address the mesh listener binds.	127.0.0.1	mesh/broker.py
SLM_MESH_SHARED_SECRET	Shared bearer token; required when the host is not local.	unset	mesh/broker.py
SLM_MESH_WS_PORT	Port used for mDNS announcement.	7900	mesh/broker.py
SLM_MESH_DISCOVERY	off disables mDNS discovery, which only applies off-host.	on	mesh/broker.py
SLM_MESH_PEER_URL	Full URL of a remote SLM instance.	unset	mesh/broker.py
SLM_MESH_TLS	1, on, true or yes uses TLS for discovered peers.	off	mesh/remote_sync.py

Provider keys are read from OPENAI_API_KEY, ANTHROPIC_API_KEY, AZURE_OPENAI_API_KEY and OPENROUTER_API_KEY when set (core/config.py provider table, llm/backbone.py).

Rate limits. The daemon starts its limiters at 30 writes and 120 reads per 60 seconds (SLM_RATE_LIMIT_WRITE, SLM_RATE_LIMIT_READ, SLM_RATE_LIMIT_WINDOW). The Settings Rate Limits group can show 100 and 300 until you save a value, because the page reads its own defaults; save a value once and both agree.

Recall depth. A recall returns up to 20 results by default (CANONICAL_RECALL_LIMIT). Callers can ask for a different number per call.

The deployment preset

config.toml may contain a [deployment] section that picks a preset. Personal is the default (no login, no PII redaction, no retention enforcement, audit on). Enterprise turns login, PII redaction and retention on. A deployment section that is present but invalid resolves to the enterprise preset, so a typo can never silently relax controls. See Security, Company Mode and Teams.

Non-English text

There is no language setting. Replace the embedder and the reranker instead, as described in Modes Explained. The reranker keys live under retrieval in config.json:

{
  "retrieval": {
    "use_cross_encoder": true,
    "cross_encoder_backend": "openai",
    "cross_encoder_endpoint": "https://models.example.test/v1/rerank",
    "cross_encoder_model": "BAAI/bge-reranker-v2-m3",
    "cross_encoder_timeout_seconds": 15.0
  }
}

The endpoint must be HTTPS off the machine (plain HTTP is allowed for loopback and, while retrieval.trust_plain_http_lan is true, numeric private-LAN addresses), may not contain user info, a query string or a fragment, and redirects are not followed. If it is unreachable, recall returns fusion-ranked results without reranking; it never silently falls back to the English model. Remote reranking sends the query and each candidate's text to that endpoint after a best-effort secret and PII filter, so only use a service you trust.

Code: core/config.py (SLMConfig, ScopeConfig, RetrievalConfig, load_deployment_config), infra/data_root.py, cli/commands.py (cmd_config), ui/js/od-settings.js, core/remote_mode.py, infra/rate_limiter.py, retrieval/remote_reranker_config.py, core/answer_check_state.py, core/memory_kind_config.py, mcp/server.py.