Org Memory

org-memory

Use the org CLI to maintain structured, linked, human-readable knowledge in org-mode files. Org files are plain text with rich structure: headlines, TODO states, tags, properties, timestamps, and links. Combined with org-roam, they form a knowledge graph backed by a SQLite database.

Shortcuts

When your human uses these patterns, act on them directly.

Org mutations

Behaviour modifiers

These prefixes change how you respond, not what you store.

Todo — create a task

t: <text> or Todo: <text> means "create a task." Extract any date or timeframe from the text and schedule it. If the text contains a relative date ("in 3 weeks", "by Friday", "next month"), compute the actual date and add --scheduled <date> or --deadline <date>.

Action: org add "$ORG_MEMORY_HUMAN_DIR/inbox.org" '<title>' --todo TODO --scheduled <date> --db "$ORG_MEMORY_HUMAN_DATABASE_LOCATION" -f json

Use --deadline instead of --scheduled when the text implies a hard due date ("by Friday", "due March 1st"). Use --scheduled for softer timing ("in 3 weeks", "next month", "tomorrow").

Examples:

• "t: submit taxes in 3 weeks" → org add .../inbox.org 'Submit taxes' --todo TODO --scheduled 2026-03-18
• "Todo: renew passport by June" → org add .../inbox.org 'Renew passport' --todo TODO --deadline 2026-06-01
• "t: call dentist tomorrow" → org add .../inbox.org 'Call dentist' --todo TODO --scheduled 2026-02-26
• "Todo: book flights" → org add .../inbox.org 'Book flights' --todo TODO (no date mentioned)

Note — for the human

r: <text> or Note: <text> means "add this to MY org files." It is always a task or reminder for the human, not for the agent.

Action: org add "$ORG_MEMORY_HUMAN_DIR/inbox.org" '<text>' --todo TODO -f json

If the note includes a date or deadline, add --scheduled <date> or --deadline <date>. If there's no date, add it without one (the human will schedule it themselves, or ask you to).

Examples:

• "r: Buy groceries" → org add .../inbox.org 'Buy groceries' --todo TODO
• "Note: Review PR #42 by Friday" → org add .../inbox.org 'Review PR #42' --todo TODO --deadline 2026-02-28
• "r: we could add feature X to the app" → org add .../inbox.org 'Add feature X to the app' --todo TODO
• "Note: send email to Donna about safeguarding" → org add .../inbox.org 'Send email to Donna about safeguarding' --todo TODO

Note vs Todo: Both create TODO headings. The difference is intent — t: / Todo: signals a concrete task (always try to extract a date), while r: / Note: is broader (ideas, reminders, observations). When there's no date, add it without one.

Edge case — ideas and observations: If the human says "r: we could do X" or "Note: idea for Y", it's still a Note. They're telling you to write it down for them. Add it as a TODO. Don't create a roam node, don't put it in the agent's knowledge base.

Schedule — reschedule a task

s: <text> means "reschedule this task." Search for the matching TODO and update its scheduled date.

Action:

1. Search: org todo list --state TODO --search '<text>' -d "$ORG_MEMORY_HUMAN_DIR" -f json
2. If exactly one match: org schedule <custom_id> <date> -d "$ORG_MEMORY_HUMAN_DIR" --db "$ORG_MEMORY_HUMAN_DATABASE_LOCATION" -f json
3. If multiple matches: show them to the human and ask which one
4. If no match: tell the human you couldn't find it

Examples:

• "s: taxes to next Friday" → find "taxes" TODO, reschedule to next Friday
• "s: dentist 2026-04-01" → find "dentist" TODO, reschedule to 2026-04-01

Done — mark complete

d: <text>, Done: <text>, or Finished: <text> means "mark this task as DONE." Search for the matching TODO and set its state.

Action:

1. Search: org todo list --state TODO --search '<text>' -d "$ORG_MEMORY_HUMAN_DIR" -f json
2. If exactly one match: org todo set <custom_id> DONE -d "$ORG_MEMORY_HUMAN_DIR" --db "$ORG_MEMORY_HUMAN_DATABASE_LOCATION" -f json
3. If multiple matches: show them to the human and ask which one
4. If no match: tell the human you couldn't find it

Examples:

• "d: pay Nigel Kerry" → find and mark DONE
• "Finished: the PR review" → find and mark DONE
• "Done: groceries" → search for "groceries", mark DONE

Know — for the agent

k: <info>, Know: <info>, or Remember: <info> means "store this in YOUR knowledge base for future recall." This is information the agent should retain across sessions.

Action: Search for an existing node first (org roam node find), then create or update.

Examples:

• "k: Sarah prefers morning meetings" → Create/update a node for Sarah in $ORG_MEMORY_AGENT_DIR
• "Remember: The API uses OAuth2, not API keys" → Create/update a node for the API in $ORG_MEMORY_AGENT_DIR

After every write — confirm

After every mutation to either directory, print a line in this exact format:

org-memory: <action> <file-path>

Examples: org-memory: added TODO to ~/org/human/inbox.org, org-memory: created node ~/org/agent/sarah.org, org-memory: updated ~/org/agent/sarah.org.

This is mandatory. Never silently write to either directory. The human should always see what you did and where.

Output format

All commands accept -f json for structured output with {"ok":true,"data":...} envelopes. Errors return {"ok":false,"error":{"type":"...","message":"..."}}. Always use -f json.

Command safety

User-provided text (task titles, note content, search terms) must be single-quoted to prevent shell expansion. Double quotes allow $(…), backticks, and variable interpolation — single quotes do not.

# Correct
org add "$ORG_MEMORY_HUMAN_DIR/inbox.org" 'User provided text' --todo TODO -f json

# Wrong — double quotes allow shell injection
org add "$ORG_MEMORY_HUMAN_DIR/inbox.org" "User provided text" --todo TODO -f json

If the text contains a literal single quote, escape it with ''':

org add "$ORG_MEMORY_HUMAN_DIR/inbox.org" 'Don'''t forget' --todo TODO -f json

For multi-line content, pipe via stdin instead of interpolating:

printf '%s' 'Long text here' | org append k4t --stdin -d "$ORG_MEMORY_AGENT_DIR" --db "$ORG_MEMORY_AGENT_DATABASE_LOCATION" -f json

Environment variable paths ($ORG_MEMORY_HUMAN_DIR, etc.) must always be double-quoted to handle spaces, but never place user text inside double quotes.

Discovery

Run org schema once to get a machine-readable description of all commands, arguments, and flags. Use this to construct valid commands without memorizing the interface.

Setup

Configuration is via environment variables. Set them in openclaw.json so they are injected into every command automatically.

Required — set these to match your directory layout:

Optional — these have sensible defaults:

Workspace dirs (*_DIR) hold tasks, inbox, and daily files. Roam dirs (*_ROAM_DIR) hold knowledge graph nodes. Databases are collocated with roam dirs by default. Roam dirs default to <workspace>/roam — roam nodes are never created in the workspace root.

If ORG_MEMORY_USE_FOR_AGENT is not true, skip the Knowledge management section. If ORG_MEMORY_USE_FOR_HUMAN is not true, skip the Task management and Batch operations sections.

Always pass --db to point at the correct database. The CLI auto-syncs the roam database after every mutation using the --db value. Without --db, the CLI defaults to <directory>/.org.db.

Initialize each enabled directory. If the directories already contain org files, sync them first:

# Sync existing files into the roam database (skip if starting fresh)
org roam sync -d "$ORG_MEMORY_AGENT_DIR" --db "$ORG_MEMORY_AGENT_DATABASE_LOCATION"
org roam sync -d "$ORG_MEMORY_HUMAN_DIR" --db "$ORG_MEMORY_HUMAN_DATABASE_LOCATION"

# Create a seed node for the agent's knowledge base (skip if files already exist)
org roam node create 'Index' -d "$ORG_MEMORY_AGENT_DIR" --db "$ORG_MEMORY_AGENT_DATABASE_LOCATION" -f json

# Build the headline index (enables CUSTOM_ID auto-assignment and file-less commands)
org index -d "$ORG_MEMORY_AGENT_DIR" --db "$ORG_MEMORY_AGENT_DATABASE_LOCATION"
org index -d "$ORG_MEMORY_HUMAN_DIR" --db "$ORG_MEMORY_HUMAN_DATABASE_LOCATION"

The roam response includes the node's ID, file path, title, and tags.

Stable identifiers (CUSTOM_ID)

Every headline created with org add is auto-assigned a short CUSTOM_ID (e.g. k4t) when an index database exists. This ID appears in the custom_id field of all JSON responses and as a column in text output.

Use CUSTOM_IDs to refer to headlines in subsequent commands — they are stable across edits and don't require a file path:

org todo set k4t DONE -d "$ORG_MEMORY_HUMAN_DIR" --db "$ORG_MEMORY_HUMAN_DATABASE_LOCATION" -f json
org schedule k4t 2026-03-15 -d "$ORG_MEMORY_HUMAN_DIR" --db "$ORG_MEMORY_HUMAN_DATABASE_LOCATION" -f json
org note k4t 'Pushed back per manager request' -d "$ORG_MEMORY_HUMAN_DIR" --db "$ORG_MEMORY_HUMAN_DATABASE_LOCATION" -f json
org append k4t 'Updated scope per review.' -d "$ORG_MEMORY_HUMAN_DIR" --db "$ORG_MEMORY_HUMAN_DATABASE_LOCATION" -f json

To backfill CUSTOM_IDs on existing headlines that don't have them:

org custom-id assign -d "$ORG_MEMORY_HUMAN_DIR" --db "$ORG_MEMORY_HUMAN_DATABASE_LOCATION"

Never address headlines by position number (pos). Positions change when files are edited — a mutation on one headline shifts the byte positions of everything after it.

Safe identifiers (in order of preference):

1. CUSTOM_ID (e.g. k4t) — stable, short, unique
2. org-id (UUID) — stable, unique
3. Exact title — stable as long as the title doesn't change

If you need to mutate multiple headlines in the same file, either:

• Use org batch for atomic multi-step operations (recommended)
• Use CUSTOM_IDs or titles, never pos
• If you must use pos, re-query after each mutation to get fresh positions

Error handling

Branch on the ok field. Handle errors by type:

• file_not_found: wrong path or deleted file
• headline_not_found: identifier doesn't match; re-query to get current state
• parse_error: file has syntax the parser can't handle; don't retry
• invalid_args: check org schema or org <command> --help

References

Read these on demand when the conversation requires them:

• Knowledge management ({baseDir}/references/knowledge-management.md): Read when ORG_MEMORY_USE_FOR_AGENT=true and you need to create/query/link roam nodes in the agent's knowledge base.
• Task management ({baseDir}/references/task-management.md): Read when ORG_MEMORY_USE_FOR_HUMAN=true and you need to query or mutate the human's tasks, use batch operations, or map natural-language queries to commands.
• Memory architecture ({baseDir}/references/memory-architecture.md): Read at session start (file structure, session routine, ambient capture guidelines). Also contains optional memory migration instructions — only follow those if the user explicitly asks to migrate from MEMORY.md to org-mode. Migration reads/writes files outside the declared directories (see reference for details).