CLI Reference

CLI Reference

The slm command has 74 top-level commands, plus slm hook, which hosts call from their hook configuration and which is not part of the normal command parser. This page groups them by task, with one short table per group. The installed help always wins over this page: slm --help lists every command, slm <command> --help lists its flags, and slm help prints a grouped overview with the topics modes, config and self-heal. Everything here was checked against the code and its --help output.

How the CLI behaves

Most commands talk to a background daemon and start it if it is not running. These commands never auto-start it, because they only read or write files: setup, mode, provider, connect, upgrade-hosts, migrate, mcp, warmup, hooks, codex, config, evolve, db, disable, enable, clear-cache, reconfigure, benchmark, rotate-token, remote, evidence, brain, diagnostics, wrap, optimize, cache, compress, help-optimize, loop, help, serve, restart, gdpr, summary, backup, models, media and features. Metadata options (--help, --version) create no data folder and start nothing.

The data folder is ~/.superlocalmemory unless SLM_DATA_DIR, SL_MEMORY_PATH or SLM_HOME says otherwise, in that order. See Installation.

Exit codes. Most commands exit 0 on success, 1 when the command ran and failed, and 2 when it refused to run (an unknown kind, a rejected --replaces, a forget with no query, a repair aimed at the wrong folder). A few commands document more: slm gdpr verify exits 0 for an intact receipt, 1 for a tampered one and 2 when the receipt is not found. slm remote check exits 1 on any failure. slm warmup exits non-zero when recall is not ready. slm media enable exits 3 when the daemon is not running and 4 when the computer has too little memory for images and documents.

Setup, configuration and status

Command	What it does	Notable flags
slm init	One-command setup: runs the wizard if SLM is not configured, installs Claude Code hooks, connects detected IDEs and warms the embedding model	--force redo setup, --auto no terminal needed (Mode A plus hooks), --gate experimental PreToolUse gate
slm setup	The interactive first-time wizard (ten steps, see Installation)	--auto use defaults, for scripts and CI; --no-laya do not install the on-device answer check (installed by default on Apple Silicon)
slm mode [a, b, c]	Show or set the operating mode; run slm restart to apply a change (Modes Explained)	--json
slm provider [set] [PROVIDER]	Show or set the language-model provider for Mode B and C. Providers: openai, anthropic, azure, ollama, openrouter, custom	--endpoint URL, --key KEY, --model NAME, --mode b or c (all for custom; --mode defaults to c)
slm models	Installed Ollama models, recommendations and the hosted catalogue. Read-only	--json
slm embedder switch, status, rollback, cancel, forget-previous, calibrate, calibration-status, calibration-cancel	Change the text embedding model in the background without stopping SLM, or calibrate the answer check on your store (Embedding Models). upgrade is still accepted but is not offered: text stays on its own model	switch MODEL: --dimension, --provider (sentence-transformers, ollama, openai or slm-media), --endpoint, --task-prefixes, --no-wait, --force; calibrate: --model (llama3.2:latest or gemma3:4b), --samples; all: --json
slm config get KEY and slm config set KEY VALUE	Read any config.json value in dot notation, or set one of a short allowed list	--json
slm status	Mode, provider, profile, data folder, database size, and whether the daemon is holding saves back	--verbose or -v adds migration log, daemon port and last version; --json
slm health	Health of the scoring, consistency and lifecycle layers (Living Brain)	--json
slm doctor	Pre-flight check of dependencies, embedding worker, daemon, database and installed hosts	--quick fast checks only, --deep read every database page, --fix repair what can be repaired first, --refile-hidden list memories that are hidden but scored to keep (with --yes, re-file them), --json
slm warmup	Load the embedding model (about 500 MB the first time) and confirm recall is ready	--timeout SECONDS (default 120; 0 checks once)
slm help [TOPIC]	Grouped overview of all commands. Topics: modes, config, self-heal	none
slm dashboard	Open the web dashboard (Dashboard Tour)	--port PORT (default 8765)

slm doctor always checks that the daemon is available, including with --quick: it reads the daemon's descriptor and makes at most one health request with a 0.5-second limit, and never starts the daemon. A missing descriptor is a warning that points at slm serve start; a recorded process that is dead, or an unhealthy or invalid daemon, is a failure that points at slm restart. The daemon check is part of the --json output too. Doctor exits 0 when every check passed or only warned and 1 when any check failed, and it prints the whole report before it exits. With Memory health on, it adds a Memory description row and a Memory repairs row (Health and Repair). Two checks to know: Failed saves warns when saves were given up on after repeated attempts and fails at 50 or more (slm enrich fix repair-saves retries them), and Answer check fit warns when the text search runs on a model the answer check is not fitted for, with slm embedder rollback as the way back (Embedding Models).

slm config set accepts only these keys and refuses the rest with a DISALLOWED_KEY error that lists them: 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 and scope.recall_include_shared. Everything else has its own command or lives in the dashboard Settings page (Configuration).

The daemon and the MCP server

Command	What it does	Notable flags
slm serve [start, stop, status, install, uninstall]	Control the background daemon. start is the default. install adds a user-level service that starts it at login: a LaunchAgent on macOS, a systemd user service on Linux, a Task Scheduler task on Windows	none
slm restart	Kill orphan processes, clear stale state, start fresh and check health. Needed after changes that cannot apply at runtime, such as a new mode	--dashboard open the dashboard afterwards, --json
slm mcp	Start the MCP server on stdio. IDE configurations run it as a subprocess; you rarely run it by hand (MCP Tools)	none
slm reap	Find orphaned SLM processes. The default only lists them	--force kill orphans, --all kill every slm mcp process (after switching IDE), --json
slm rotate-token	Rotate the install token. Run slm restart afterwards	none
slm token show	Print the SuperLocalMemory key, readable only by you	none

Memory operations

Command	What it does	Notable flags
slm remember CONTENT	Store a memory. The default receipt returns once the memory is searchable by its wording; enrichment continues in the background (How Remember Works)	--tags a,b, --kind KIND, --replaces ID, --sync wait for completion, --scope personal, shared or global, --shared-with PROFILES, --json
slm recall QUERY (alias slm search)	Multi-channel retrieval (How Recall Works)	See the recall flags below
slm trace QUERY	Recall with a per-channel score breakdown	--limit N (default 10), --json
slm list	Recent memories, newest first, with each one's kind and the id that update and delete take	-n or --limit N (default 20), --kind, --tag (repeatable), --tags-match all or any, --json
slm update FACT_ID CONTENT	Replace the text of one memory. The corrected fact is re-indexed everywhere. A review-gated change returns a correction case id	--json
slm delete FACT_ID	Delete one memory by exact id	--yes or -y, --json
slm outcome success, partial or failure FACT_ID...	Tell SLM how using recalled memories went; it is the command-line twin of the report_outcome tool (Active Memory)	--query-id ID (from slm recall --json), --note TEXT, --agent NAME, --report-id KEY (a retry with it records nothing new), --json
slm forget QUERY	Delete every memory whose text contains the query (case-insensitive), after a confirmation	--dry-run preview only (with no query it previews everything), --yes or -y, --json
slm observe [CONTENT]	Offer content for automatic capture; SLM decides whether it holds a decision, bug fix or preference worth storing. Reads a pipe or an argument	none

--kind takes one of the nine memory kinds: rule, decision, status, procedure, prospective, opinion, correction, episodic or semantic. --replaces names an earlier memory this one supersedes; the old one stops being returned, and review_correction undoes it.

The forget naming collision. The CLI command slm forget QUERY deletes memories whose text matches the query. The MCP tool forget does something different: it runs the Ebbinghaus decay cycle, which moves memories between lifecycle zones and is a dry run by default. It deletes nothing by query. Use slm delete or the MCP delete_memory tool to remove a single memory by id.

The recall flags narrow a search; filters that match nothing are reported rather than ignored.

Flag	Meaning
--limit N	Maximum results (default 20)
--window SPAN	Event-time range: a relative span (24h, 7d, 30d, 1y) or 2026-07-01..2026-07-31
--project P, --project-strict, --prefer-project P	Only that project's memories (not narrowed if none match, unless strict), or rank them higher without removing others
--saved-by AGENT, --about NAME, --kind KIND	Only memories saved by an agent, mentioning a name, or of a kind
--tag LABEL (repeatable), --tags-match all or any	Only memories with the tag or tags
--as-of T, --known-as-of T, --valid-at T, --include-unknown	Point-in-time recall (ISO-8601); --include-unknown adds memories whose temporal provenance is unknown
--include-global, --no-global, --include-shared, --no-shared	Include or exclude other profiles' global and shared memories (Profiles)
--depth 0, 1 or 2	How much of each memory to return: 0 a glance (ids and one-line previews), 1 the full memory (default), 2 with its source memory, same-session neighbours, memories about the same entities and the correction that replaced it
--max-tokens N	Cap the returned results at about this many tokens; never exceeded
--show-weak	When nothing supports the question, show the closest weak matches labelled [weak] instead of nothing
--fast	Skip the internal agentic verification round
--json	Structured output

The CLI shows related memories, labelled [related], when nothing answers the question exactly, and prints "That isn't in SuperLocalMemory." when nothing is close. While SLM is still warming up it prints "SuperLocalMemory is still warming up — results may be incomplete. Try again in a few minutes." instead of that. With Answer Check running, recall output also carries one line saying whether the top results answer the question (How Recall Works).

Kinds, summaries, saved views and corrections

Command	What it does	Notable flags
slm kinds status	Counts per kind, the backend in use, and any classification run	--json
slm kinds settings	Show or change memory-kind settings	--enable or --disable, --backend auto, rules, laya, jev, llm or off, --jev-consent yes or no, --standing-rules on or off
slm kinds backfill start, pause, resume, cancel, revert, status	Classify existing memories as a background, undoable run (revert RUN_ID)	--json
slm kinds set FACT_ID KIND	Set and confirm one memory's kind	--json
slm kinds review and slm kinds confirm	List suggestions waiting for confirmation, and confirm 1 to 200 at once (FACT_ID or FACT_ID=KIND)	review: --kind, --limit
slm summary session ID, sessions, day [DATE], project [PATH]	Readable, traceable summaries; extractive unless a model is configured. No daemon needed	--profile P, --json
slm view list, create NAME QUERY, run NAME, show NAME, rename NAME NEW, delete NAME	Saved views: named recall queries you can re-run. Deleting a view touches no memory	create: --kind, --window, --as-of, --limit (1 to 50)
slm review-correction CASE_ID apply, reject or rollback EXPECTED_VERSION	Resolve a review-gated correction case; the version is a compare-and-swap guard	--event-valid-until T (apply only), --json
slm corrections overtaken and slm corrections restore-overtaken CASE_ID	List correction cases that your own later delete, replace or edit closed, and put one back	--limit (1 to 500), --profile, --json

See Memory Kinds for what the nine kinds mean and how suggestions are confirmed.

Pictures, documents, folders and other sources

Command	What it does	Notable flags
slm features	What is on and what you can turn on: images and documents, folder sources, bot messages	--json
slm media enable, disable, status, gc, repair	Turn images and documents on or off, check set-up, find leftovers, and make pictures findable by what they show (Pictures and Documents)	enable: --yes; disable: --remove-files; gc: --apply; repair: --dry-run; all: --json
slm sources add PATH, list, report ID, rescan ID, remove ID, forget-empty ID	Connect folders and Obsidian vaults. Sources are read, never written	add: --kind folder or obsidian, --yes; remove: --purge also erases the memories it gave, --yes; forget-empty: --yes
slm adapters list, enable, disable, start, stop, status [NAME]	Manage ingestion adapters: gmail, calendar, transcript (Ingestion Adapters)	none
slm ingest	Import external observations into SLM learning	--source ecc or jsonl, --file FILE, --dry-run, --json

Profiles

Command	What it does	Notable flags
slm profile list, switch NAME, create NAME	List memory profiles, change the active one, or create one. Deleting a profile is a dashboard action, not a CLI command (Profiles)	--json

Data, integrity, repair, backup and restore

Command	What it does	Notable flags
slm db migrate	Apply additive, forward-only schema migrations. They also apply automatically at startup	--status inspect, --dry-run preview
slm db integrity	Read-only store check: pages, relations, fidelity, projections, repair status. Safe while SLM runs (Health and Repair)	--pages read every page (slow), --json
slm db repair	Fix orphan rows, stale or orphan vectors, erased-text leftovers and settled obligations, with receipts. Preview by default	--apply, --undo RUN_ID, --root FOLDER (required to change anything), --batch-size, --pause-ms, --max-seconds, --json
slm db repair-entities	Undo wrong entity links left by the older fuzzy name merge. Preview by default; applying takes a backup first and writes an undo record	--apply, --undo RUN_ID, --root FOLDER (required to change), --json
slm db fidelity	List facts that no longer say what their memory said (a number, a date, a dropped "never")	--profile, --limit, --withhold FACT_ID and --release FACT_ID (SLM must be stopped), --json
slm db regraph	Re-derive the graph copy of your memories from the store	--check report drift only, --profile
slm db reembed	Backfill embeddings for facts that never got one	--all-profiles, --limit N, --json
slm db compact	Drop old LanceDB vector-store versions. The background job already keeps only an hour of history (SLM_VECTOR_HISTORY_DAYS changes it); use this with SLM stopped and --offline to shrink a store that already grew	--offline (daemon stopped; also removes unverified files)
slm db scale status, adopt, prepare, verify, promote, rollback	Stage and promote the optional CozoDB and LanceDB projections; SQLite stays canonical	--stage-id ID (verify, promote), --backup-id ID (rollback)
slm db restore-points	List the copies of your memories you can go back to	--json
slm db restore POINT_ID	Go back to a restore point, applied when SLM next starts (Host Upgrades)	--yes, --cancel, --no-reimport, --json
slm db prepare-downgrade	Get your memories ready to run on the previous SLM version	--yes, --cancel, --json
slm migrate	A one-shot migrator for data from a very old install; a current install does not need it. Not the same as slm db migrate	--rollback only while the migrator's own backup still exists
slm evidence export DEST, verify BUNDLE, import BUNDLE, rebuild	Versioned, checksummed JSONL evidence bundles. import and rebuild are dry runs unless --execute	export: --profile; import: --profile, --replace, --rollback-dir, --execute; all: --json
slm backup status, recovery-key, decrypt FILE	Cloud-backup encryption: state, the recovery key, and turning a downloaded backup into a plain .db file (Cloud Backup)	recovery-key --import reads a key from stdin; decrypt: -o OUT, --recovery-key-stdin, --force; all: --json
slm backup restore-safety-copy	Put back the safety copy that a failed dashboard restore kept. The service must be stopped	--id POINT restore a different recovery backup, --json
slm backup clear-recovery-needed	Let the service start again after you repaired the stores yourself. Deletes the pre-restore copy the marker kept	--json
slm clear-cache	Wipe regenerable caches. memory.db and learning.db are kept	none
slm enrich status, pause RUN, resume RUN, undo [RUN], fix REPAIR	Memory health: the automatic description of older memories, the repairs that wait for your say-so (entity-links, refile-hidden, repair-saves) and undo (Health and Repair)	undo and fix: --yes (needed without a terminal); fix: --include-undone; all: --json

Restore points are a command-line and API feature; the dashboard has no restore-point screen. The dashboard restores a different thing, a recovery backup, from Backup & Cloud (Cloud Backup).

Governance, GDPR and teams

Command	What it does	Notable flags
slm gdpr status	Compliance posture: receipts, audit events, known gaps. Read-only	--profile, --json
slm gdpr export	Subject-access and portability export for one profile	--profile NAME (required), --output FILE (default stdout), --json
slm gdpr erase	Irreversibly erase a profile. Needs both --profile and --yes; preview with --dry-run	--profile, --yes, --dry-run, --json
slm gdpr verify	Check the HMAC of an erasure receipt (exit 0 intact, 1 tampered, 2 not found)	--receipt-id ID, --profile, --json
slm team status	Whether login is required and how many users exist	--json
slm team policy	Require login for this workspace, or stop requiring it	--require-login on or off

slm gdpr offers the controls a GDPR programme needs; whether a deployment is compliant depends on how it is configured and operated. See Governance and Security, Company Mode and Teams. slm team runs on the SLM computer and acts as the break-glass path for company mode.

Mesh and remote access

Command	What it does	Notable flags
slm mesh status and slm mesh peers	Local agent-mesh broker health and statistics, or the active peer sessions (Multi-Agent Memory, Bots That Talk)	none
slm remote tls init	Create a certificate authority and a server certificate for remote access	--name HOST and --ip ADDR (repeatable), --days N (at most 397), --force
slm remote enable and slm remote disable	Turn the TLS listener on or off; takes effect on restart	enable --listen HOST:PORT
slm remote keys add NAME	Create a key, shown once, bound to exactly one profile	--read-only recall only, --profile P (default: the active profile)
slm remote keys list, allow KEY mesh or media, disallow KEY mesh or media, revoke KEY	List keys (never their secrets), grant or remove the optional mesh and media abilities, or revoke a key on its next request	list: --json
slm remote check	Check remote access; exits 1 on any failure	--json

slm remote serves other computers on your network and is separate from Web Access and Connected Apps, which connects AI apps on the internet from the dashboard. A key's holder can read the full text of every memory in its profile, so create keys sparingly.

Optimize

Command	What it does	Notable flags
slm optimize status, on, off, savings	Show or switch the Optimize cache and compression settings, and report token and cost savings (Optimize)	savings: --since DAYS (default 7), --provider anthropic, openai or gemini; all: --json
slm cache status, clear, invalidate, ttl, semantic	Cache state, clearing, tag invalidation, TTLs, and the semantic cache switch	--tenant T (default default), invalidate --tag TAG, ttl --set SECONDS --semantic SECONDS, semantic on or off, --json
slm compress status, mode, prose, code, ccr, align	Compression state and mode (safe or aggressive); prose on or off. code, ccr and align still parse but only print a notice	--json
slm proxy	Start the optimization proxy	--port (default 8765), --provider anthropic, openai or gemini, --no-compress, --semantic, --json
slm wrap AGENT [ARGS]	Point an agent at the proxy and launch it	--list registered agents, --dry-run, --persistent
slm help-optimize [TOPIC]	Full Optimize reference. Topics: cache, compress, optimize, proxy, agents, safety	--no-pager

Hooks and IDE integration

Command	What it does	Notable flags
slm connect [IDE]	Write SLM's MCP entry into an IDE's config (IDE Setup)	--list, --here project scope, --dry-run, --transport stdio, http or http-mcp-remote, --port (default 8765), --verify, --profile NAME (writes SLM_MCP_PROFILE), --cross-platform, --disable ADAPTER, --json
slm upgrade-hosts	Preview, then refresh, installed host integrations after an SLM upgrade. Read-only unless --apply (Host Upgrades)	--host NAME (repeatable), --all-detected, --apply
slm hooks install, remove, status	Manage additive auto-capture hooks for Claude Code or Codex. status is the default	--agent claude or codex, --gate, --dry-run
slm codex install, remove, status	Manage the Codex skills, subagents and lifecycle hooks	--dry-run
slm hook EVENT	Run one hook handler. Hosts call it; you do not	none

slm hook takes one event name and bypasses the normal parser so it starts fast. The events are mandate, start, gate, init-done, checkpoint, stop, user_prompt_submit, post_tool_async, post_tool_outcome, user_prompt_rehash, stop_outcome, auto_recall, topic_shift, before_web, and the Codex events codex-start, codex-prompt and codex-stop. An unknown event prints an error and exits 1.

Sessions and lifecycle

Command	What it does	Notable flags
slm session open	Warm a session's context	--project-path, --query, --max-results (default 10)
slm session close	Close a session and create its temporal summaries	--session-id (default: the most recent)
slm session-context [QUERY]	Print relevant memories for hooks. Uses a fast SQLite path	--max-age-days N (default 30; 0 turns the age filter off), --full, --json
slm decay	Ebbinghaus forgetting cycle. Previews by default	--execute apply, --dry-run, --profile, --json
slm quantize	Embedding-precision cycle. Previews by default	--execute, --dry-run, --profile, --repair-sketches, --json
slm consolidate	The consolidation pipeline	--cognitive also run cognitive consolidation, --dry-run, --profile, --json
slm soft-prompts	List the auto-learned patterns that become soft prompts	--profile, --json

slm quantize previews by default. --repair-sketches repairs only the compressed search sketches (the QJL ones) that slm doctor can recommend repairing: it rebuilds legacy or incompatible sketches from the original vectors that are still kept, and changes neither the stored vectors, nor the base compressed bytes, nor bit widths, nor lifecycle state. It starts no embedding model, a repeat skips sketches that are already compatible, and memories without a usable original vector are reported as needing to be embedded again. It reuses the existing rotation; if the rotation matrix is missing or incompatible, restore it from a backup before applying. The full --execute cycle also adjusts precision according to retention, and before it applies anything it prints counts for sketch repairs, downgrades, upgrades and memories moving to the forgotten state (the vectors are kept). The counts can change if writes happen meanwhile. Both forms accept --profile and --json; with --json the preview goes to the error stream and the final JSON carries it under data.plan.

The cycles and what they do are explained in Living Brain.

Bounded loops, evolution and the brain

Command	What it does	Notable flags
slm loop demo	Run the keyless convergence demo: a stub proposer and a gate that fails twice and passes on lap 3. A good post-upgrade check (Bounded Loops)	--iterations N (default 10), --json
slm loop history	List recorded runs for a loop name	--name NAME (default convergence-demo), --json
slm loop show RUN_ID	Show every lap of one run	--json
slm evolve	Post-session skill evolution. The Stop hook normally runs it (Skill Evolution)	--session ID, --profile ID
slm brain [status]	Read-only, profile-scoped Living Brain evidence summary	--json

Running a loop against a real gate is done with the MCP tools slm_loop_run, slm_loop_history and slm_loop_show; the CLI has only demo, history and show.

Diagnostics

Command	What it does	Notable flags
slm diagnostics export DEST	Write a deterministic, content-free JSON report of local operational aggregates, for support	--json
slm diagnostics reliability	Ask whether wired mechanisms really work: has each learner moved off its prior, has each guarded path ever run	--min-observations N
slm ops list, resolve OPERATION_ID, status	List failed, stuck or degraded operations; retry, force-reconcile or cancel one (administrator action); quick failure count	list: --profile; resolve: --action retry, force_reconcile or cancel; all: --json
slm benchmark	Run the evo-memory benchmark against an isolated temporary database. It never touches your data	--json

Escape hatches

Command	What it does	Notable flags
slm disable	Switch SLM off globally: writes ~/.superlocalmemory/.disabled and stops the daemon	--reason TEXT
slm enable	Remove the .disabled marker and print the command that starts the daemon	none
slm reconfigure	Re-run the interactive post-install configurator (profile and install-time knobs)	extra flags pass through

slm clear-cache, slm reap and slm restart are also escape hatches and are listed above. The environment variable SLM_DISABLE=1 has the same effect as slm disable for one shell.

Agent-native JSON output

Commands that list --json in their help print one compact JSON object on a single line. The envelope is the same everywhere: success, command, version, then data on success or error on failure, with optional metadata and next_actions that name follow-up commands.

{
  "success": true,
  "command": "recall",
  "version": "<installed-version>",
  "data": {
    "results": [
      {
        "fact_id": "abc123",
        "content": "Database uses PostgreSQL 16",
        "relevance_score": 0.87,
        "ranking_score": 0.0132,
        "memory_confidence": 0.7,
        "rank_position": 1
      }
    ],
    "count": 1,
    "score_contract_version": "2",
    "calibration_status": "uncalibrated",
    "answer_confidence": null,
    "answer_check_status": "off"
  },
  "next_actions": [
    {"command": "slm list --json", "description": "List recent memories"}
  ]
}

Recall keeps ranking relevance separate from stored-memory confidence (Retrieval Score Contract). The full recall response also carries query_id, no_confident_match, abstained, abstention_reason, answerability, channel_status, incomplete_channels, project_scope, tag_scope, reranker_status and the answer-check fields described in Answer Check. When no_confident_match is true or answer_confidence is low, the CLI help advises an agent to rewrite the query into one to three more specific sub-queries and recall again before concluding nothing was found.

slm recall "auth" --json | jq '.data.results[0].content'
slm list --json | jq '.data.results[].fact_id'
slm status --json | jq '.data.mode'

In CI, slm status --json | jq -e '.success' is a quick health gate, and slm remember "Deployed $SHA" --json records a deployment.

CLI or MCP

Need	Use	Example
IDE integration	MCP, in the profile that fits (MCP Tools)	slm connect cursor --profile code
Shell scripts and CI	CLI with --json	slm recall "auth" --json
Agent frameworks	CLI with --json, or the adapters in Framework Adapters	slm remember "..." --json
Humans	CLI, readable output	slm recall "auth"

Code: cli/main.py (parser and dispatch), cli/commands.py, cli/json_output.py, cli/kinds_cmd.py, cli/summary_cmd.py, cli/view_cmd.py, cli/media_cmd.py, cli/sources_cmd.py, cli/remote_commands.py, cli/embedder_cmd.py, cli/integrity_cmd.py, cli/gdpr_cmd.py, cli/service_installer.py, hooks/hook_handlers.py.