Install memex for OpenClaw

You are inside an OpenClaw agent. The user asked you to set up memex — the local-first MCP memory server — so that:

1. Every OpenClaw session in ~/.openclaw/agents/main/sessions/ is auto-captured into ~/.memex/data/memex.db (SQLite + FTS5).
2. Existing past sessions are back-filled in one shot.
3. The OpenClaw gateway gains the memex MCP server with 11 search/retrieval tools (memex_search, memex_recent, memex_overview, memex_get_conversation, etc.).
4. Capture continues automatically thereafter (systemd user-service on Linux, LaunchAgent on macOS).

Where you might be running:

• 🖥️ VPS (Linux VPS hosting OpenClaw + maybe Hermes) — typical "always-on agent" deployment
• 💻 macOS workstation running OpenClaw alongside Claude Code / etc.
• 💻 Linux workstation (Pop!OS, Fedora, etc.) running OpenClaw locally
• 🤖 Moonshot Kimi-Claw VPS — also works (memex-mvp ≥ 0.11.5 auto-detects the merged-file Kimi-Claw layout)

All work the same way. Platform (Linux/macOS) determines daemon type. Deployment kind (self-hosted vs Kimi-Claw vs workstation) is auto-detected by the memex-mvp daemon.

The contract: zero questions to the user. Discovery → actions → verification. Each step prints a single line of result. The final message tells the user exactly what to do next.

⚠ For ClawHub reviewers, security scanners, and curious users

This skill installs memex-mvp from npm (if not already present) and wires it as an MCP server inside the OpenClaw gateway config. memex-mvp is the canonical install path for OpenClaw — the older "memex-openclaw OpenClaw-native plugin" approach (npm package @parallelclaw/memex-openclaw) is deprecated as of 2026-05-27; see deprecation notes in that package's README. The daemon path works on every OpenClaw deployment without security-scanner friction.

Hard guarantees:

• Every command is printed to the user before running it
• sudo is NEVER invoked by the skill (linger fallback prints the manual command for the user)
• OpenClaw config is merged, never overwritten — other plugins / MCP servers preserved
• memex at runtime emits zero outbound network traffic — local-first by design
• If memex-mvp is already installed (e.g. via the generic install-memex skill), this skill skips re-install and only wires OpenClaw — no surprise reinstalls

Migration note for users who tried the deprecated memex-openclaw plugin

If the user previously installed @parallelclaw/memex-openclaw via openclaw plugins install, that plugin is now deprecated. The daemon-based approach below handles the same use case more reliably (no OpenClaw plugin security-scanner friction, no per-version manifest changes, no allowConversationAccess opt-in). To clean up before installing daemon-based:

# 1. Uninstall the old plugin (no-op if not installed)
openclaw plugins uninstall memex-openclaw 2>&1 || true

# 2. Restore openclaw.json from the .bak the plugin left (if you want to undo the plugin's edits)
[ -f ~/.openclaw/openclaw.json.bak ] && echo "found backup, you can restore via:" \
  && echo "  cp ~/.openclaw/openclaw.json.bak ~/.openclaw/openclaw.json"

# 3. Restart gateway so the old plugin stops loading
openclaw gateway restart

The data the old plugin wrote (raw_type='openclaw-agent-end' rows) stays in memex.db and remains searchable — the daemon's writes will appear alongside under the same source. UNIQUE(source, conversation_id, msg_id) prevents duplicates.

Before this skill can be installed (one-time OpenClaw bootstrap)

If the user runs openclaw skill install install-memex-claw and gets back "plugin not allowed" / "permission denied" / "skill plugin disabled" — the gateway's permission model is blocking arbitrary skill execution by default. This is OpenClaw's safety guard, not specific to memex. Check your OpenClaw docs for the exact permission name (common: plugins.allow: ["skill", …] in ~/.openclaw/openclaw.json or skills.enabled: true), set it once, restart the gateway, and retry.

Prerequisite

You are running INSIDE an OpenClaw agent that can execute shell commands. If you're not OpenClaw, stop: this is the wrong skill — use install-memex instead.

Step 1 — Discovery (run ALL checks first, then decide)

Run each as a separate shell command, capture the output. Don't act on anything until all checks have answered.

# 1. Are we inside OpenClaw?
which openclaw 2>/dev/null || find / -maxdepth 4 -name "openclaw" -type d 2>/dev/null | head -3

# 2. Platform — Linux or Darwin (macOS)
uname -s

# 3. Node version (need ≥ 22.12 for memex-mvp ≥ 0.11.x)
node --version

# 4. Existing memex install?
which memex && memex --version 2>&1 || echo "NO_MEMEX"

# 5. Existing memex daemon running?
memex-sync status 2>/dev/null | head -5 || echo "NO_DAEMON"

# 6. Existing OpenClaw sessions?
ls -1 ~/.openclaw/agents/main/sessions/ 2>/dev/null \
  | grep -E '^[0-9a-f-]+\.jsonl$' | wc -l

# 7. On Linux: systemd user-instance available?
[ "$(uname -s)" = "Linux" ] && systemctl --user --version 2>/dev/null | head -1 \
  || echo "macOS or no-systemd"

# 8. OpenClaw gateway config
ls -la ~/.openclaw/openclaw.json 2>&1

# 9. Was the deprecated memex-openclaw plugin previously installed?
ls -la ~/.openclaw/npm/node_modules/@parallelclaw/memex-openclaw 2>/dev/null \
  || echo "no old plugin"

Branch on results:

Report each result back as facts, not questions:

> "✓ OpenClaw detected at /usr/local/bin/openclaw"

> "✓ Linux + systemd 255 — will use user-systemd for daemon"

> "✓ Node 22.22.2 — ok"

> "→ memex not installed — installing 0.11.6+"

> "✓ Found 60 OpenClaw sessions ready to back-fill"

> "⚠ Deprecated memex-openclaw plugin detected — will uninstall first"

Step 2 — Install memex-mvp (skip if already at ≥ 0.11.6)

npm install -g memex-mvp@latest

If EACCES (no permission to write to npm global prefix), apply the standard fix — never use sudo:

mkdir -p ~/.npm-global
npm config set prefix ~/.npm-global
echo 'export PATH=$HOME/.npm-global/bin:$PATH' >> ~/.zshrc 2>/dev/null
echo 'export PATH=$HOME/.npm-global/bin:$PATH' >> ~/.bashrc 2>/dev/null
export PATH=$HOME/.npm-global/bin:$PATH
npm install -g memex-mvp@latest

Verify:

memex --version    # must print 0.11.6 or later

Step 3 — Install the auto-capture daemon (skip if already running)

memex-sync install

Platform-specific behavior (handled automatically by memex-sync):

If loginctl enable-linger needs sudo, the install output prints the exact command. Tell the user that line verbatim — don't try to sudo yourself.

Verify daemon running:

memex-sync status
# Should show: process: running (pid XXXX)

Step 4 — Back-fill existing sessions

memex-sync scan

memex-sync scan walks the source dirs (Claude Code on workstations, OpenClaw on VPS, etc.), reads each .jsonl, filters internal-state noise (.checkpoint., .trajectory., .reset.* snapshots — they'd double-count), and writes to memex.db.

Expected output:

=== Claude Code + Cowork ===
- skipping claude-code: directory not found at /home/user/.claude/projects   (on VPS without Claude Code)
- skipping claude-cowork: directory not found at ...                          (on Linux — Cowork is macOS-only)
scanning openclaw: /home/user/.openclaw/agents/main/sessions
+ openclaw-abc12345.jsonl ← 23 msgs from openclaw (ai-titled)
+ openclaw-def67890.jsonl ← 41 msgs from openclaw
…
scanned 60 files · 1316 messages emitted

If the daemon was already running (Step 3 skipped), scan is still useful — it catches anything the daemon's chokidar may have raced on startup.

Step 4b — Channel-aware re-import (recommended once after first install)

memex-mvp v0.11.5+ does two-mode auto-detection (self-hosted OpenClaw with separate .jsonl per session vs Moonshot Kimi-Claw with merged-file layout) and channel routing (Telegram chat / Kimi-web session / openclaw-cli). After the first scan, run:

memex-sync backfill-channels --yes

This wipes existing source = 'openclaw' rows and re-imports with the current channel-aware pipeline. Use it once after upgrading from any memex-mvp earlier than 0.11.5, and any time the user reports "messages from one chat are merged into another" symptoms.

Step 5 — Wire memex as an MCP server in openclaw.json (REQUIRED for LLM retrieval)

The daemon handles capture; the memex-mvp MCP server exposes memex_search, memex_recent, memex_overview, etc. to the OpenClaw LLM toolset. Without this step, the LLM has no way to search memex from inside a conversation — it would fall back to raw sqlite3 shell calls (or fail).

Config-key gotcha

OpenClaw 2026.5+ reads MCP servers from cfg.mcp.servers (nested), NOT from the top-level cfg.mcpServers (flat). Verify with openclaw mcp list after the merge — must show memex.

Find the absolute path to memex (MCP stdio doesn't inherit shell PATH):

which memex

Merge memex into ~/.openclaw/openclaw.json at mcp.servers.memex (preserves other servers; idempotent):

python3 <<'PY'
import json, shutil
from pathlib import Path
p = Path.home() / ".openclaw" / "openclaw.json"
cfg = json.loads(p.read_text())
memex_bin = shutil.which("memex")
if not memex_bin:
    print("ERROR: memex binary not in PATH; rerun Step 2")
    raise SystemExit(1)

# Correct path is mcp.servers (nested), NOT mcpServers (flat).
mcp_section = cfg.setdefault("mcp", {})
servers = mcp_section.setdefault("servers", {})
servers["memex"] = {"command": memex_bin, "args": [], "env": {}}

# Clean up any legacy top-level mcpServers.memex from earlier skill versions
stale = cfg.get("mcpServers", {})
if isinstance(stale, dict) and "memex" in stale:
    del stale["memex"]
    if not stale:
        cfg.pop("mcpServers", None)
    print("cleaned stale mcpServers.memex from previous skill version")

p.write_text(json.dumps(cfg, indent=2))
print(f"memex MCP wired at {p} → {memex_bin}")
PY

Verify the merge worked (THREE checks):

# 1. Config has the right key
python3 -c "
import json
from pathlib import Path
cfg = json.loads((Path.home() / '.openclaw' / 'openclaw.json').read_text())
s = cfg.get('mcp', {}).get('servers', {}).get('memex')
print('mcp.servers.memex:', json.dumps(s, indent=2) if s else 'MISSING')
"

# 2. OpenClaw's own listing recognises it (authoritative check)
openclaw mcp list 2>&1 | grep -i memex
# Must show 'memex' or similar.

# 3. (after restart, see Step 7d) LLM toolset includes memex_search

Step 6 — Restart OpenClaw

openclaw gateway restart

Wait ~5 seconds, then check logs:

journalctl --user -u openclaw -n 50 --no-pager 2>/dev/null | grep -iE "memex|mcp" | tail -10 \
  || tail -100 ~/.openclaw/logs/gateway.log 2>/dev/null | grep -iE "memex|mcp" | tail -10

You should see the gateway picking up the memex MCP server.

Step 7 — Verify (four checks — all must pass)

7a. Daemon is alive and capturing

memex-sync status
# process: running (pid XXXX)

# Watch the daemon log for a few seconds
tail -10 ~/.memex/logs/memex-sync.log 2>/dev/null

7b. memex.db has OpenClaw rows

sqlite3 ~/.memex/data/memex.db <<'SQL'
.headers on
.mode column
SELECT COUNT(*) AS total_openclaw_rows,
       MIN(ts) AS earliest_ts,
       MAX(ts) AS latest_ts
  FROM messages
 WHERE source = 'openclaw';
SQL

Expected: total_openclaw_rows ≥ 1 (much more if you ran back-fill on existing sessions).

7c. Live capture writes a NEW row

PRE=$(sqlite3 ~/.memex/data/memex.db \
  "SELECT COUNT(*) FROM messages WHERE source='openclaw'")
echo "PRE: $PRE"

# >>> NOW send a test message in OpenClaw, wait for reply <<<
sleep 8

POST=$(sqlite3 ~/.memex/data/memex.db \
  "SELECT COUNT(*) FROM messages WHERE source='openclaw'")
echo "POST: $POST  diff=$((POST - PRE))"

sqlite3 ~/.memex/data/memex.db \
  "SELECT id, role, datetime(ts,'unixepoch','localtime') AS when_,
          substr(text,1,60) AS preview, conversation_id
     FROM messages WHERE source='openclaw' ORDER BY ts DESC LIMIT 5"

Expected: diff ≥ 2 (user + assistant), recent rows include your test message.

7d. LLM toolset includes memex tools (requires a FRESH OpenClaw session)

MCP tools are loaded into the LLM's function-calling toolset at session start. The session in which you just ran openclaw gateway restart has the OLD toolset. To verify memex is exposed:

1. Start a fresh OpenClaw conversation.
2. Ask the agent: "Find the earliest conversations from April using memex_search."
3. The agent should:
• Mention memex_search in its reasoning or just call it
• Return real results (IDs, timestamps, snippets from ~/.memex/data/memex.db)
• NOT fall back to direct sqlite3 calls

If the agent doesn't see memex tools → re-check Step 5 (probably mcp.servers vs mcpServers key issue) and openclaw mcp list.

Step 8 — Final message to user

After all checks pass:

✅ memex wired into OpenClaw.

Captured:
  • Every new OpenClaw turn writes to ~/.memex/data/memex.db within seconds
    (the memex-sync daemon watches ~/.openclaw/agents/main/sessions/)
  • Past sessions back-filled (Step 4)
  • Channel routing works for self-hosted OpenClaw + Moonshot Kimi-Claw

Search:
  • From inside OpenClaw chat: ask the model to search its memory — uses memex_search
  • From CLI: `memex search "query"` or `memex overview`
  • From any other MCP-enabled client on the same machine (Claude Code, Cursor, etc.) —
    the memex MCP server exposes 11 search tools

Daemon:
  • Status check: `memex-sync status`
  • Auto-restarts on login (LaunchAgent on macOS, systemd-user on Linux)
  • Survives SSH logout if linger is enabled

Pair with:
  • install-memex (generic skill) — if you also use Claude Code, Cursor, Cline, etc.
    on this same machine, they all share ~/.memex/data/memex.db
  • memex-hermes (pip package) — if you run Hermes Agent on this machine

What this skill explicitly does NOT do

• ❌ Install Node.js (manual step if needed — print instructions, don't auto-install)
• ❌ Use sudo. Ever. If permissions block something, surface the manual fix and stop.
• ❌ Install the deprecated @parallelclaw/memex-openclaw plugin. The daemon path supersedes it.
• ❌ Overwrite existing mcp.servers entries. Merge.
• ❌ Configure outbound network access. memex is local-first.

Edge cases

Safety rules — read before starting

• Never sudo. Never edit system files. Never modify $PATH outside ~/.zshrc / ~/.bashrc with the user's consent.
• Always confirm before deleting anything from the user's machine.
• Always preserve existing config keys when editing JSON.
• Print every shell command to the user before running it.
• If any step fails, stop. Don't try to recover by trying random things. Print the error and ask the user.

Begin

Run Step 1 (Discovery) now. Print results. Branch as the table indicates.