Agent Memory Setup v2 (Gemini Embeddings 2)

Create a 3-tier memory directory structure (HOT/WARM/COLD) for OpenClaw agents and configure the built-in memory-core plugin to use Google Gemini Embeddings...

Security Scan

The skill does what it says: it creates local memory directories and gives configuration steps to enable Google Gemini embeddings; it asks for a Gemini API key only and contains no hidden network calls or unrelated credential requests.

Purpose & Capability

Name/description match the delivered artifacts: a small setup script that creates HOT/WARM/COLD memory files and SKILL.md instructions to configure OpenClaw to use Gemini embeddings. No unrelated binaries, installs, or credentials are requested.

Instruction Scope

SKILL.md instructs only to run the provided bash script (which only writes local files), edit openclaw.json, export GEMINI_API_KEY, and restart the gateway. It warns that subsequent embedding calls will send memory content to Google's API—no instructions read unrelated system files or exfiltrate data.

Install Mechanism

No install spec; the skill is instruction-only plus a single local script. The script is a small, readable bash file and does not download or execute remote code.

Credentials

The only credential the skill requires in practice is the GEMINI_API_KEY described in the README; that is appropriate for enabling a cloud embedding provider. No other secrets or unrelated environment variables are requested.

Persistence & Privilege

always:false and the skill does not alter other skills or system-wide config automatically. It needs a manual edit of openclaw.json and a manual gateway restart; normal autonomous invocation settings are unchanged.

Assessment

This skill is small and coherent, but before installing consider: (1) After setup, your agent will send memory file contents to Google Gemini for embedding — do not store sensitive secrets or personal data in these memory files unless you accept that external transmission. (2) Use a dedicated Google API key/project to limit blast radius and billing exposure; treat GEMINI_API_KEY as secret and rotate/revoke if needed. (3) Review and commit the openclaw.json changes yourself (the skill instructs you to edit it) rather than blindly applying them. (4) The optional 'lossless-claw' plugin is a third-party component — inspect it before installing. (5) The provided script only writes local files to the workspace you specify; run it against a test workspace first if you want to verify behavior.

Create a 3-tier memory directory structure for OpenClaw agents and configure semantic search using Google Gemini Embeddings 2.

What This Skill Does

1. Creates directory structure and stub files via a bash script (no network calls, no env reads, no dependencies)
2. Provides configuration instructions for openclaw.json to enable Gemini-based memory search

Privacy Notice

⚠️ After setup, the agent's memory_search tool sends memory file content to Google's Gemini embedding API for vectorization. This is how semantic search works — files must be embedded to be searchable. The setup script itself makes no external calls.

Prerequisite

Google Gemini API key — free at https://aistudio.google.com/apikey

Setup

Step 1: Create directory structure

bash scripts/setup_memory_v2.sh /path/to/agent/workspace

Creates: memory/, memory/hot/, memory/warm/, stub .md files, heartbeat-state.json.

Step 2: Configure openclaw.json

Add under agents.defaults:

"memorySearch": { "provider": "gemini" },
"compaction": { "mode": "safeguard" },
"contextPruning": { "mode": "cache-ttl", "ttl": "1h" },
"heartbeat": { "every": "1h" }

Set API key: export GEMINI_API_KEY=your-key

Enable plugin: "lossless-claw": { "enabled": true }

Step 3: Restart

openclaw gateway restart

Memory Tiers

• 🔥 HOT (memory/hot/HOT_MEMORY.md) — Active session state, pending actions
• 🌡️ WARM (memory/warm/WARM_MEMORY.md) — Stable preferences, references
• ❄️ COLD (MEMORY.md) — Long-term milestones and distilled lessons

Optional Plugin

Lossless Claw (@martian-engineering/lossless-claw) — compacts old context into expandable summaries to prevent amnesia. Install separately: openclaw plugins install @martian-engineering/lossless-claw