CLI Reference

CLI Reference

The recollectium command is the main user-facing interface to Recollectium Core. Commands print human-readable output by default when they return command data. Use --json or cli_output: json when scripts and adapters need machine-readable output. Use --compact or --verbose to choose how much response detail a command returns. Non-argparse runtime failures follow the same output preference on stderr.

Global form:

recollectium [GLOBAL OPTIONS] COMMAND [COMMAND OPTIONS]

Global options

Global options normally go before the command. --json and --human-readable choose rendering. --compact and --verbose choose response detail. The rendering flags may appear before or after the command.

Option	Values	What it does
--config CONFIG_PATH	path	Uses a specific JSON config file instead of the default XDG config path. Explicit missing paths fail except for config creation commands.
--log-level LEVEL	debug, info, warning, error	Overrides the configured log level for this run only. It does not write to the config file.
--json	none	Prints JSON success and failure output for this invocation, overriding cli_output. Mutually exclusive with --human-readable.
--human-readable	none	Prints human-readable success and non-argparse failure output for this invocation, overriding cli_output. Mutually exclusive with --json.
--compact	none	Prints compact response payloads for this invocation, overriding response_verbosity. Works with either rendering mode.
--verbose	none	Prints full response payloads for this invocation, overriding response_verbosity. Works with either rendering mode.
--version	none	Prints the installed Recollectium version and exits.

Example:

recollectium --log-level debug list
recollectium list --memory-space project-alpha
recollectium list --json
recollectium service discover --json

init

recollectium init

Initializes Recollectium for first use. It creates the config file if needed, creates XDG directories, creates or migrates the SQLite database, prepares the built-in embedding model cache, and writes model readiness metadata.

This command uses the configured database folder and default memory space. It does not accept a database-path override.

Notes:

• First run may download the FastEmbed model cache and can take longer than later runs. If existing memories need the active embedding profile, init refreshes them inline.
• Success prints initialized config, data, cache, logs, runtime, database, and embedding model paths.

add

recollectium add --space user --memory-space project-alpha --type preference --content "I prefer concise answers."
recollectium add --space workspace --workspace-uid my-project --memory-space project-alpha --type decision --content "Use SQLite for local storage."

Adds one memory and stores an embedding for semantic search. Use --memory-space to route the write to a non-default memory space.

Options:

Option	Required	Values	What it does
	--space SPACE	Yes	user, workspace
	--memory-space KEY	No	Routes the write to a specific configured memory space.
--type TYPE	Yes	user or workspace memory type	Places the memory in a bucket. User types include fact, preference, personal_fact, social_context, goal, communication_style, note. Workspace types include fact, decision, task_context, configuration, bug_finding, note.
--content CONTENT	Yes	text	The memory text to store and embed.
--workspace-uid UID	Required for workspace, forbidden for user	string	Stable workspace identifier. Use the repo name or project folder name when possible.
--metadata JSON_OR_PATH	No	JSON object or @path/to/file.json	Optional structured metadata. Must be a JSON object.
--source SOURCE	No	string	Optional label for where the memory came from, such as chat, manual, issue, or adapter.
--confidence SCORE	No	number from 0.0 to 1.0	Optional confidence score.
--sensitivity LABEL	No	string	Optional privacy or sensitivity label for future privacy-aware handling.

search-user

recollectium search-user --memory-space project-alpha [--type TYPE] [--limit LIMIT] [--protected-minimum N] [--match-threshold VALUE] [--include-archived] "query"

Searches user-space memories by semantic similarity. This is the right command when the question is about the person using the agent: preferences, communication style, personal facts, durable goals, and cross-project context.

Arguments and options:

Name	Required	Default	What it does
query	Yes	none	Natural language search text. Exact wording is not required because search is semantic.
--memory-space KEY	No	Routes the search to a specific configured memory space.
--type TYPE	No	all user types	Narrows results to one user memory bucket. Usually skip this first and add it only if results are too broad.
--limit LIMIT	No	20	Maximum number of ranked results. Must be positive.
--protected-minimum N	No	3	Keeps this many top-ranked results before threshold filtering applies. Must be an integer >= 0.
--match-threshold VALUE	No	model_recommended_default	Threshold for results after the protected minimum. Use model_recommended_default, null, or a number between 0.0 and 1.0.
--include-archived	No	false	Includes archived memories in the search candidate set.

search-workspace

recollectium search-workspace --workspace-uid UID --memory-space project-alpha [--type TYPE] [--limit LIMIT] [--protected-minimum N] [--match-threshold VALUE] [--include-archived] "query"

Searches memories for one workspace. This is the right command when the question is about a repo, project, implementation detail, task, decision, bug, or configuration.

Arguments and options:

Name	Required	Default	What it does
query	Yes	none	Natural language search text.
--workspace-uid UID	Yes	none	Workspace to search. Aliases resolve to the canonical UID.
--memory-space KEY	No	Routes the search to a specific configured memory space.
--type TYPE	No	all workspace types	Narrows results to one workspace bucket, such as decision or configuration.
--limit LIMIT	No	20	Maximum number of ranked results. Must be positive.
--protected-minimum N	No	3	Keeps this many top-ranked results before threshold filtering applies. Must be an integer >= 0.
--match-threshold VALUE	No	model_recommended_default	Threshold for results after the protected minimum. Use model_recommended_default, null, or a number between 0.0 and 1.0.
--include-archived	No	false	Includes archived workspace memories.

list

recollectium list [--memory-space KEY] [--space SPACE] [--type TYPE] [--status STATUS] [--workspace-uid UID] [--include-archived] [--limit LIMIT]

Lists memories directly rather than semantically searching them. Use it for inspection, audits, debugging, or finding IDs.

Options:

Option	Required	What it does
--memory-space KEY	No	Routes the list to a specific configured memory space.
--space SPACE	No	Filters to user or workspace.
--type TYPE	No	Filters to one memory type.
--status STATUS	No	Filters by status, commonly active or archived.
--workspace-uid UID	No	Filters workspace memories to one workspace UID.
--include-archived	No	Includes archived memories. Without it, archived memories are hidden.
--limit LIMIT	No	Maximum number of memories to return. Must be positive.

get

recollectium get MEMORY_ID [--memory-space KEY]

Retrieves one memory by ID and prints it as JSON. Use list or search-user / search-workspace first if you do not know the ID.

Arguments:

Argument	Required	What it does
MEMORY_ID	Yes	The memory ID to fetch.
--memory-space KEY	No	Routes the lookup to a specific configured memory space.

update

recollectium update MEMORY_ID [--memory-space KEY] [--type TYPE] [--content CONTENT] [--metadata JSON_OR_PATH] [--source SOURCE] [--confidence SCORE] [--sensitivity LABEL]

Updates editable fields on one memory. Updating --content regenerates that memory's embedding.

Arguments and options:

Name	Required	What it does
MEMORY_ID	Yes	Memory to update.
--memory-space KEY	No	Routes the update to a specific configured memory space.
--type TYPE	No	Replaces the memory type bucket.
--content CONTENT	No	Replaces the memory text and regenerates the embedding.
--metadata JSON_OR_PATH	No	Replaces metadata with a JSON object, either inline or from @path/to/file.json.
--source SOURCE	No	Replaces the source label.
--confidence SCORE	No	Replaces the confidence score. Must be between 0.0 and 1.0.
--sensitivity LABEL	No	Replaces the sensitivity label.

Use recollectium upgrade for package upgrades. recollectium update is only for memory records.

archive

recollectium archive MEMORY_ID [--memory-space KEY]

Marks a memory as archived. Archived memories are not hard-deleted, but they are hidden from normal list and search results unless you pass --include-archived to the relevant command.

Arguments:

Argument	Required	What it does
MEMORY_ID	Yes	Memory to archive.
--memory-space KEY	No	Routes the archive to a specific configured memory space.

db-status

recollectium db-status [--memory-space KEY]

Shows database schema migration status as JSON for the selected memory space. It initializes the database if needed and reports current and pending schema versions.

Use --memory-space KEY before the command to inspect a non-default memory space.

dev

recollectium dev true
recollectium dev reset
recollectium dev eval
recollectium dev optimize-threshold [--start N] [--end N] [--step N] [--beta N] [--format png|csv] [--output PATH] [--write-config]
recollectium dev false

recollectium dev true routes future Recollectium activity to the seeded development database. The fixture contains 100 user memories across 10 topic buckets and 90 workspace memories across 3 workspaces. Each workspace has three 10-memory themes for retrieval evaluation, stored in dev_theme metadata.

recollectium dev reset recreates the configured seeded database from the canonical fixture. It targets development.seeded_database_path and does not touch the normal memory-space database.

recollectium dev eval runs a seeded retrieval-quality check against the development fixture. It reports five separate metrics: Exact MRR, Semantic MRR, Thematic Weighted Precision@10, Thematic Weighted Recall@10, and Ranked-set NDCG@5. Human-readable compact output shows the metrics summary, while --verbose adds the seeded database path, regular database path, preparation steps, and diagnostics. In a human-readable TTY, it may also show live phase and status updates on stderr, while the final summary stays on stdout and JSON output remains structured and automation-safe. The eval refuses to run if the seeded database path matches the selected regular database path.

recollectium dev optimize-threshold runs the seeded thematic label threshold optimizer. It loads the seeded development database, sweeps thresholds from --start to --end by --step, and recommends a retrieval.match_threshold by maximizing weighted F-beta over the labeled candidate pools. With no beta specified, the default is --beta 0.5 for F0.5, which biases toward precision and fewer noisy matches. --beta 1.0 is the balanced precision/recall option, and higher beta values favor recall. In human-readable output, --compact shows the recommendation and metrics, while --verbose also shows the threshold sweep range. The command writes a PNG graph by default, or CSV with --format csv; CSV to stdout keeps progress and summary text on stderr so pipes stay clean. The summary includes the recommendation, weighted precision, weighted recall, F-beta score, confuser exposure, unrelated exposure, and average returned count. It is advisory unless --write-config is passed.

Examples:

recollectium dev optimize-threshold --format png --output threshold-sweep.png
recollectium dev optimize-threshold --beta 0.5 --format csv --output precision-sweep.csv
recollectium dev optimize-threshold --beta 2.0 --write-config

recollectium dev false returns future activity to the regular user database without deleting either database.

The dev switch and reset commands refuse to run while a managed Recollectium API or MCP service is running. Stop or restart the service around the switch so service traffic cannot keep using the previously opened database.

webui

recollectium webui serve [--host HOST] [--port PORT]
recollectium webui start
recollectium webui stop
recollectium webui restart
recollectium webui status

The WebUI is the local browser control plane for Core. It mirrors the service lifecycle shape used by the API and MCP services, but targets browser use instead of client integration.

webui serve runs the browser control plane in the foreground for development. webui start, webui stop, webui restart, and webui status manage the background WebUI service using the configured host and port.

Default host and port are 127.0.0.1 and 8766. Keep the WebUI bound to localhost unless private-network controls protect the endpoint.

workspace

Workspace commands manage workspace UIDs, aliases, and renames.

workspace list

recollectium workspace list [--include-archived] [--include-aliases]

Lists known workspace UIDs.

Option	Required	What it does
--include-archived	No	Includes UIDs that only appear on archived memories.
--include-aliases	No	Returns workspace objects with nested alias arrays instead of plain UID strings.

workspace resolve

recollectium workspace resolve UID

Normalizes a workspace UID candidate and resolves aliases to the canonical workspace UID.

workspace rename

recollectium workspace rename OLD_UID NEW_UID

Migrates all memories from OLD_UID to NEW_UID, including archived memories. Both values are normalized according to workspace.uid_normalization before the operation.

workspace alias add

recollectium workspace alias add CANONICAL_UID ALIAS_UID [--migrate-existing]

Creates an alias so an old or alternate workspace name resolves to a canonical workspace UID.

Option	Required	What it does
--migrate-existing	No	Moves existing memories under the alias UID to the canonical UID in the same transaction. Use this when the alias name already has memories.

workspace alias list

recollectium workspace alias list [--workspace UID]

Lists alias mappings. With --workspace, filters to one canonical workspace UID.

workspace alias remove

recollectium workspace alias remove ALIAS_UID

Removes an alias mapping. It does not delete memories.

config

recollectium config
recollectium config --path
recollectium config --defaults
recollectium config --validate
recollectium config get KEY
recollectium config set KEY VALUE
recollectium config unset KEY
recollectium config init [--force]
recollectium config doctor
recollectium config edit
recollectium config reset

Configuration commands inspect and modify the JSON config file.

Command	What it does
config	Prints the effective config: built-in defaults merged with the config file.
config --path	Prints the resolved config file path without creating a file.
config --defaults	Prints built-in default values without creating a file.
config --validate	Validates the active config. Exits 0 if valid, 1 if missing, 2 if invalid.
config get KEY	Prints one effective config value by dot-notation key, such as service.port.
config set KEY VALUE	Writes a config value. VALUE is parsed as JSON when possible, otherwise stored as a string.
config unset KEY	Removes an explicit key so the built-in default applies again.
config init --force	Creates a starter config file. --force overwrites an existing file.
config doctor	Checks config validity and filesystem writability for data, cache, logs, runtime, and database parent directories.
config edit	Opens the config file in $EDITOR, creating it first if needed.
config reset	Replaces the config file with built-in defaults.

Supported keys are listed in Configuration.

service

recollectium service start api
recollectium service start mcp
recollectium service stop
recollectium service status
recollectium service discover
recollectium service restart [--type api|mcp]

Managed services run in the background and write PID and discovery files under the runtime directory.

Command	What it does
service start api	Starts the local HTTP JSON API service.
service start mcp	Starts the managed HTTP MCP service.
service stop	Stops the running managed Recollectium service, if one exists.
service status	Prints whether a managed service is running, plus type, PID, and endpoint when available.
service discover	Prints human-readable discovery output by default. Use service discover --json for machine-readable adapter parsing. Exits 1 with status: "not_running" when no service is running.
service restart	Restarts the currently running service or the last known service type.
service restart --type api	Restarts API service when no running or previous service type can be inferred.
service restart --type mcp	Same, but for MCP service.

dev serve

recollectium dev serve [--host HOST] [--port PORT]

Runs the local HTTP API service in the foreground for development and debugging. For managed background use, prefer recollectium service start api.

Options:

Option	Required	Default	What it does
--host HOST	No	service.host, usually 127.0.0.1	Bind interface. Non-local binds can expose unauthenticated memory operations.
--port PORT	No	service.port, usually 8765	TCP port for the API service.

mcp-stdio

recollectium mcp-stdio

Runs an MCP server over stdin/stdout. This is intended for MCP clients that spawn Recollectium as a child process. The server runs for the lifetime of the client connection and does not create a PID file.

Use --config and --log-level before the command if needed.

embedding commands

recollectium embedding-status [--memory-space KEY]
recollectium embedding-maintenance [--memory-space KEY]
recollectium embedding-jobs [--memory-space KEY] [--job-id JOB_ID] [--state STATE] [--limit LIMIT]
recollectium embedding-refresh [--memory-space KEY] [--space SPACE] [--workspace-uid UID] [--include-archived]
recollectium embedding-jobs-clear [--memory-space KEY] [--state STATE]... --yes

embedding-status

Prints the active local embedding profile, runtime status, startup re-embedding job reference, and recent job metadata. Use --memory-space when you want the status for a non-default memory space.

embedding-maintenance

Prepares the configured embedding model and refreshes stale or missing embeddings inline. Installers and successful package upgrades run this automatically; users can also run it after changing embedding settings or when troubleshooting model/cache state. Use --memory-space to target a non-default memory space.

embedding-jobs

Lists embedding jobs or fetches one job by ID.

Option	Required	What it does
--memory-space KEY	No	Routes the job lookup or listing to a specific configured memory space.
--job-id JOB_ID	No	Fetches exactly one embedding job. Without it, the command lists jobs.
--state STATE	No	Filters list mode by state, such as pending, in_progress, completed, or failed.
--limit LIMIT	No	Limits list mode to a positive number of jobs.

Embedding jobs are used for model readiness and re-embedding memories when the active embedding profile changes or stale embeddings are detected. Refresh jobs run inline, so the command that triggers refresh waits for completion or returns a failure with the job status path.

embedding-refresh

Forces stale memories to be re-embedded for the active embedding profile. It processes the refresh inline and returns after the job completes or fails.

Option	Required	What it does
--memory-space KEY	No	Routes the refresh to a specific configured memory space.
--space SPACE	No	Limits refresh to user or workspace memories.
--workspace-uid UID	No	Limits refresh to one workspace UID. Implies workspace scope.
--include-archived	No	Includes archived memories in the refresh candidate set.

If no stale memories match, the command reports refreshed: false, stale_count: 0, and no job.

embedding-jobs-clear

Deletes embedding job audit records without deleting memories or embeddings.

Option	Required	What it does
--memory-space KEY	No	Routes the cleanup to a specific configured memory space.
--state STATE	No	Deletes matching job states. Repeat for multiple states. If omitted, deletes completed, failed, and pending.
--yes	Yes	Confirms deletion of job records. Required so scripts do not clear job audit records accidentally.

lifecycle commands

completion

recollectium completion [SHELL]
recollectium completion [SHELL] --source
recollectium completion [SHELL] --install [--yes]

Supported shells are bash, zsh, fish, and powershell. If SHELL is omitted, Recollectium tries to detect it.

Option	Required	What it does
SHELL	No	Shell to configure.
--source	No	Prints only the raw completion script for eval or shell startup files.
--install	No	Writes a managed completion block into the shell startup file. Prompts before writing.
--yes	No	Skips the confirmation prompt when used with --install.

upgrade

recollectium upgrade [--check] [--dry-run] [--force] [--version VERSION | --main] [--install-method METHOD] [--repo OWNER/REPO] [--allow-main] [--timeout SECONDS]

Checks for a newer release and upgrades using the detected install method. Without --version or --main, follows the install metadata target. New installs track the latest release unless pinned to a version or main.

Option	Required	Default	What it does
--check	No	false	Checks for an available upgrade and prints the plan without applying it.
--dry-run	No	false	Prints the command that would run without applying it.
--force	No	false	Builds and applies an upgrade plan even if versions appear current.
--version VERSION	No	recorded target	Targets a release. Use latest to track latest releases, or a version/tag such as 1.0.2 or v1.0.2 to pin that release. Mutually exclusive with --main.
--main	No	recorded target	Tracks and installs from the main branch. Mutually exclusive with --version.
--install-method METHOD	No	auto	Overrides detection. Values: auto, bootstrap, pip, pipx, uv_tool, source.
--repo OWNER/REPO	No	Recollectium repo	Checks a different GitHub repo. Passing this also permits main fallback when no release exists.
--allow-main	No	false	Permits main-branch fallback for bootstrap or source upgrades if no release exists.
--timeout SECONDS	No	600	Package-manager command timeout.

uninstall

recollectium uninstall [--purge] [--yes-delete-all-recollectium-data] [--dry-run]

Prints safe uninstall instructions and stops managed services. By default, it preserves memories and config, but removes the Recollectium-owned derived local model cache and other derived artifacts. It may edit shell profile or startup files only to remove Recollectium-managed completion blocks, not arbitrary shell configuration. Source and unknown installs report a manual package-removal hint because safe self-removal is not supported for those install types.

Option	Required	What it does
--purge	No	Permanently deletes Recollectium-owned config, data, cache, logs, runtime paths, and memories after confirmation.
--yes-delete-all-recollectium-data	Required only for non-interactive purge	Confirms destructive deletion. Requires --purge.
--dry-run	No	Shows planned actions without stopping services or deleting files. With --purge, previews deletion paths.

In human-readable mode, --compact prints a short uninstall sentence and --verbose prints the full result payload. Both can show live progress on stderr while package removal runs when the terminal supports it. JSON output stays structured and automation-safe.

Exit behavior and output

• Human-readable output is the default. When the target output stream is a TTY, human-readable output uses Rich-backed ANSI color for headings, field labels, errors, and hints. Set cli_output to json in config, or pass --json, when scripts or adapters need machine-readable output.
• Pass --json to force JSON for one invocation, even when config prefers human-readable output. This is the safest mode for scripts and adapters.
• --json and --human-readable are mutually exclusive and may appear before or after the command. If you need one of those strings as a literal command value, put it after --.
• --compact and --verbose choose response detail. They work with either rendering mode and override response_verbosity for one invocation.
• Protocol commands keep their machine contract: completion --source, completion candidate generation, dev serve, and mcp-stdio do not switch to human text.
• Non-argparse failures follow the selected output format on stderr.
• Validation errors usually exit 2.
• Runtime, service, database, migration, not-found, and embedding errors usually exit 1.
• service discover intentionally prints status: "not_running" to stdout and exits 1 when no managed service is running, so adapters can read the state. Use --json for adapter or script discovery calls.