clipmem 0.1.2

macOS clipboard memory backed by SQLite and searchable from OpenClaw
Documentation

clipmem

Crates.io License: MIT Rust 1.87+

A searchable, local-only clipboard history for macOS, with a JSON-first CLI designed so agents — OpenClaw and others — can recall things you've copied.

clipmem watches the macOS system clipboard (NSPasteboard), archives every observed state into a local SQLite database, and exposes a handful of retrieval commands (recall, recent, timeline, search, get, export) that return either human-readable text or structured JSON.

Contents

Requirements

  • macOS. Clipboard capture uses NSPasteboard via objc2 / objc2-app-kit and is gated behind cfg(target_os = "macos"). The database and search layers compile on other Unixes for development, but capture will not.
  • Apple Silicon for the prebuilt Homebrew package (target aarch64-apple-darwin). Intel Macs are supported via cargo install.
  • Rust 1.87+ for building from source (rust-version in Cargo.toml).

Install

Homebrew (Apple Silicon):

brew install tristanmanchester/tap/clipmem

Cargo (Apple Silicon or Intel):

cargo install clipmem

Build from source:

cargo build --release
# or install the current checkout into ~/.local/bin
cargo install --path . --root ~/.local --force --locked

All three paths produce the same clipmem binary.

Quick start

Capture the current clipboard once to confirm everything works:

clipmem capture-once

Start the watcher in the foreground (useful for testing — see Run in the background for the normal setup):

clipmem watch --interval-ms 350

Recall what you copied, or list recent items:

clipmem recall "what was that shell command?"
clipmem recent --hours 24

Check SQLite / FTS5 diagnostics if something looks off:

clipmem doctor

Run in the background (LaunchAgent)

The normal way to use clipmem is as a user LaunchAgent that starts on login. The included script installs the binary, creates the database directory with tight permissions, and loads the agent:

./scripts/install_launchagent.sh

By default that will:

  • install the Rust binary into ~/.local/bin
  • create ~/Library/Application Support/clipmem (mode 0700)
  • write ~/Library/LaunchAgents/io.openclaw.clipmem.watch.plist
  • launch the watcher with --skip-initial so the pasteboard state present at login isn't re-logged
  • kickstart the agent immediately

Verify it's running:

launchctl list | grep clipmem
# expect: <pid>  0  io.openclaw.clipmem.watch
tail -f ~/Library/Application\ Support/clipmem/logs/clipmem.stderr.log

Environment variables the install script honors:

  • CLIPMEM_INSTALL_ROOT — defaults to ~/.local
  • CLIPMEM_DB_PATH — defaults to ~/Library/Application Support/clipmem/clipmem.sqlite3
  • CLIPMEM_INTERVAL_MS — defaults to 350

To remove the LaunchAgent and plist (leaves the database intact):

./scripts/uninstall_launchagent.sh

For a full wipe including the database, see Uninstall / Full cleanup.

Recalling and searching

Start with the command that returns the least structure needed for the job:

  • clipmem recall — best-first answer plus a few alternatives. The primary agent-facing retrieval command.
  • clipmem recent — recent unique clipboard states, deduplicated by snapshot.
  • clipmem timeline — chronological capture events, including repeated copies of the same content.
  • clipmem search — direct lexical matching over stored text.
  • clipmem get <snapshot-id> — nested item/representation detail for a single snapshot.
  • clipmem export <snapshot-id> --item <n> --uti <uti> --out <path> — raw bytes for binary payloads.

Common recipes

# Best-first answer
clipmem recall "what was that git one-liner?"
clipmem recall --prefer-recent --hours 24 --format json
clipmem recall "Terminal stuff" --prefer-app terminal --format toon

# Recent unique items from one app
clipmem recent --hours 24 --app safari --format md

# Paginated chronological history
clipmem timeline --hours 24 --limit 25 --format json
clipmem timeline --hours 24 --limit 25 --cursor "<next_cursor>" --format json

# Lexical search, forcing a mode
clipmem search "launchctl bootstrap"
clipmem search --mode literal "50%"
clipmem search --mode fts "\"launchctl\" AND bootstrap"

# Detail + raw bytes
clipmem get 42 --format json
clipmem export 42 --item 0 --uti public.png --out ./clipboard.png

Shared retrieval filters

search, recent, timeline, and recall all accept the same filters. get and export accept them as guards against the explicitly targeted snapshot.

Time:

  • --since <RFC3339> — captures at or after this timestamp (e.g. 2026-04-16T09:00:00Z).
  • --until <RFC3339> — captures at or before this timestamp.
  • --hours <N> — last N hours, unless --since is also provided (then --since wins).

Source:

  • --app <name> — case-insensitive substring match on the recorded frontmost app name.
  • --bundle-id <id> — case-insensitive exact match on bundle identifier, e.g. com.apple.Safari.

Content shape:

  • --kind text|html|rtf|url|file|image|pdf|binary|other (file means file URLs; other means mixed or empty snapshots).
  • --has-text, --has-url, --has-file-url, --has-image, --has-pdf — presence flags are additive with AND semantics.

Size:

  • --min-bytes <N> / --max-bytes <N>.

Pagination

Every list command accepts --limit (bounded 1–250, default 10) plus an opaque --cursor returned as next_cursor in a prior response:

clipmem search "git status" --format json --limit 10
clipmem search "git status" --format json --limit 10 --cursor "<next_cursor>"

Cursors are opaque and tied to the active query, mode, and filters.

Search modes

search and recall accept --mode auto|fts|literal. Auto mode picks FTS or literal per query — you rarely need to override. It prefers literal matching for URLs, paths, bundle ids, dotted identifiers, and shell-like fragments (more reliable for punctuation-heavy clipboard data); plain-text queries can still use FTS first.

recall extras

On top of the shared retrieval filters, recall supports:

  • --format md|json|toon (default: md)
  • --limit — ranked candidates to consider (default 5)
  • --full — expand the best candidate text instead of the compact form
  • --quote — force quoted best-text output
  • --min-score <0.0-1.0> — minimum normalized match score before a query stands on its own
  • --prefer-recent — bias ranking toward recency
  • --prefer-app <name> — bias toward matching app or bundle id
  • --hours <N> — window for the recent fallback when a query is weak

Output formats

Human by default, machine-readable on request:

# Human-readable terminal output (default for search/recent/timeline/get)
clipmem recent --hours 24

# Machine-readable JSON for scripts and agents
clipmem recent --hours 24 --format json

Supported formats, by command:

  • search, recent, timeline, get: text (default), json, jsonl, md, toon (not toon for get, which returns nested detail).
  • recall: md (default), json, toon.
  • capture-once and doctor: --json only (alias for structured output).
  • export: writes raw bytes to --out; no --format.

--json on search, recent, timeline, get, capture-once, and doctor is a compatibility alias for --format json.

JSON envelope

search, recent, timeline, and recall return a stable top-level envelope:

  • schema_version
  • command
  • generated_at
  • applied_filters
  • truncated
  • next_cursor
  • results

Flattened text fields

get --format json and retrieval rows surface flat text fields so agents don't have to walk the nested representation tree:

  • best_text, best_text_uti
  • text_fragments
  • urls, file_paths
  • html_text, rtf_text
  • text_summary

search and recall additionally include:

  • why_matched
  • matched_fields

The full nested items[].representations[] structure is still present on get for deep inspection and export workflows.

Script-friendly by design

  • stdout contains only the requested command output
  • stderr contains diagnostics only
  • no interactive prompts anywhere
  • list commands use bounded defaults and opaque cursor pagination

Exit codes:

  • 0 — success
  • 2 — invalid args
  • 3 — not found
  • 4 — unsupported format
  • 5 — database error
  • 6 — platform error
  • 1 — uncategorized runtime failure

Using with OpenClaw

Install the packaged skill via the binary rather than by copying files:

clipmem agents openclaw install-skill

By default this installs into the active OpenClaw workspace:

~/.openclaw/workspace/skills/clipboard-memory

The workspace root is resolved from openclaw config get agents.defaults.workspace, falling back to ~/.openclaw/workspace.

To install into the shared OpenClaw skills directory instead:

clipmem agents openclaw install-skill --shared
# writes to ~/.openclaw/skills/clipboard-memory

Or to a specific destination:

clipmem agents openclaw install-skill --dest /path/to/skill --force

Other useful commands:

clipmem agents openclaw doctor         # check PATH, workspace, installed files, sandbox
clipmem agents openclaw print-skill    # print the packaged SKILL.md
clipmem agents openclaw uninstall-skill

What the skill ships

SKILL.md
references/commands.md
references/troubleshooting.md

The packaged skill points OpenClaw at the JSON-first retrieval commands:

  • clipmem recall "<query>" --format json
  • clipmem recall --prefer-recent --hours 24 --format json
  • clipmem timeline ... --format json when chronology matters
  • clipmem search ... --format json when direct lexical matching matters
  • clipmem get <snapshot-id> --format json when deeper nested detail is needed

Troubleshooting

OpenClaw must be able to run clipmem from its own environment, not just from your interactive shell. If you installed clipmem into ~/.local/bin, make sure that directory is on the PATH seen by OpenClaw. If sandboxing is active, the binary may also need to be available inside the sandbox image or container.

Start with:

clipmem agents openclaw doctor
openclaw sandbox explain   # when available

For empty results, sandbox visibility, or binary-only (image / PDF) snapshots where exact text isn't available, see extras/openclaw/clipboard-memory/references/troubleshooting.md.

Example OpenClaw prompts once installed

  • "Find that ffmpeg command I copied yesterday."
  • "Search my clipboard history for the SQL migration with WAL mode."
  • "What was the URL I copied from Safari about objc2 NSPasteboard?"
  • "Show me the full clipboard entry for snapshot 128."

Skill packaging (reference)

  • skills/clipboard-memory/ — the canonical cross-agent skill source that skills.sh discovers from the repo root.
  • extras/openclaw/clipboard-memory/ — the OpenClaw-native package installed by clipmem agents openclaw install-skill.
  • extras/agent-skills/clipboard-memory/ — a portable packaging mirror for generic agent-skill runtimes.

For compatibility, ./scripts/install_openclaw_skill.sh still exists as a thin wrapper around clipmem agents openclaw install-skill --shared --force.

Privacy

  • All clipboard data stays on your machine. clipmem makes no network calls and sends no telemetry.
  • The database lives at ~/Library/Application Support/clipmem/clipmem.sqlite3. Logs are in the same directory under logs/.
  • Permissions: directory 0700, database and logs 0600 — enforced at runtime in src/db.rs::harden_path_permissions and by the install script's umask 077.
  • The database is not encrypted. Rely on FileVault (or similar disk encryption) for at-rest protection.
  • There is no automatic retention or pruning. The archive grows until you delete it. See Uninstall / Full cleanup to wipe.

Uninstall / Full cleanup

# stop and remove the LaunchAgent
./scripts/uninstall_launchagent.sh

# remove the OpenClaw skill (if installed)
clipmem agents openclaw uninstall-skill

# delete the database and logs
rm -rf ~/Library/Application\ Support/clipmem

# remove the binary
cargo uninstall clipmem || rm -f ~/.local/bin/clipmem

# if installed via Homebrew
brew uninstall clipmem 2>/dev/null || true

Limitations

  • Text, HTML, URLs, file URLs, RTF, JSON, and XML are indexed when a reasonable text form is available. Images, PDFs, and opaque binary types are stored as blobs but are not OCR'd.
  • The watcher polls NSPasteboard.changeCount on a short interval. It is best-effort — if the clipboard changes multiple times within the poll window (400ms by default), intermediate states can be missed.
  • The frontmost app is recorded as a practical hint, not a guaranteed pasteboard owner.
  • RTF and HTML text extraction is intentionally lightweight.
  • Search is great for commands, code, URLs, notes, logs, and copied prose. It is not semantic search. --mode auto is the default; use --mode fts for strict FTS5 queries or --mode literal for exact substring matching.
  • clipmem get --format json omits raw blob bytes (flattened text fields are included). Use clipmem export to recover binary payloads.
  • --format toon is only supported for flattened output (search, recent, timeline, recall), not for nested get detail.

How it works

For each observed clipboard state, clipmem stores:

  • the whole clipboard snapshot
  • each pasteboard item inside that snapshot
  • every representation type the item exposes (identified by UTI — Uniform Type Identifier, e.g. public.png, public.url)
  • raw bytes for every representation
  • decoded text when the representation is text-like, plus strict UTF-8/UTF-16 recovery for some byte-only payloads
  • a searchable text projection that powers default retrieval
  • the frontmost application at capture time as a best-effort source hint

Identical clipboard snapshots are deduplicated by SHA-256 fingerprint, so copying the same thing ten times does not create ten full blob copies — it creates one snapshot and ten capture_events.

Default search uses SQLite's built-in full-text search (FTS5) when that fits the query, and falls back to literal substring matching for wildcard-like input, invalid FTS syntax, or zero FTS hits.

Schema

  • snapshots — deduplicated clipboard states
  • snapshot_items — items inside a snapshot
  • item_representations — one row per item/type pair, with raw blob storage
  • capture_events — each time a snapshot was observed
  • snapshots_fts — FTS5 external-content index over snapshots.search_text

Development

Project layout:

  • src/ — Rust source. Capture is gated behind cfg(target_os = "macos"); database, search, and tests compile cross-platform.
  • extras/launchd/ — LaunchAgent plist template.
  • skills/clipboard-memory/ — canonical public skill content for repo-based installers.
  • extras/openclaw/clipboard-memory/ — packaged OpenClaw-native skill content.
  • extras/agent-skills/clipboard-memory/ — portable skill package mirror for non-OpenClaw runtimes.
  • scripts/install_launchagent.sh / scripts/uninstall_launchagent.sh — install and remove the LaunchAgent.
  • scripts/install_openclaw_skill.sh — compatibility wrapper around clipmem agents openclaw install-skill.

Build and test (Rust 1.87+):

cargo build --release
cargo test

The code is written to make extension straightforward — export commands, embeddings, OCR, richer source-app heuristics, or fuller HTML parsing all fit the existing shape.

See RELEASING.md for the release workflow.

License

clipmem is released under the MIT License. See LICENSE.