Integrate OpenClaw with the Long-Term Memory Service

Step 1: Install OpenClaw and the Memory plugin

1. Install OpenClaw globally if it is not already installed.

   npm install -g openclaw@latest

2. Install the ADBPG long-term memory plugin.

   openclaw plugins install @adbpg-ai/openclaw-mem0

3. Verify that the plugin is recognized.

   openclaw plugins list | grep openclaw-mem0

   The output contains the Memory (Mem0) — ADBPG Build entry with the status enabled.

openclaw plugins uninstall openclaw-mem0
openclaw plugins install @adbpg-ai/openclaw-mem0

Step 2: Configure openclaw.json

Edit the configuration file ~/.openclaw/openclaw.json and add the openclaw-mem0 plugin configuration.

{
  "plugins": {
    "allow": ["openclaw-mem0"],
    "slots": {
      "memory": "openclaw-mem0"
    },
    "entries": {
      "openclaw-mem0": {
        "enabled": true,
        "config": {
          "mode": "platform",
          "apiKey": "${ADBPG_API_KEY}",
          "baseUrl": "${ADBPG_URL}",
          "userId": "alice"
        },
        "hooks": {
          "allowConversationAccess": true
        }
      }
    }
  }
}

Configuration parameters

Restart the gateway for the configuration to take effect:

openclaw gateway restart

Step 3: Verify basic memory operations

The mem0 CLI subcommand reads from and writes to the long-term memory service.

1. Run a health check.

   openclaw mem0 status

2. Store a preference.

   openclaw mem0 add "I like to drink a cup of black coffee every morning" --user-id alice

   Expected output:

   Added 1 memory(s):
     :  [ADD]

3. Perform a semantic search.

   openclaw mem0 search "beverage preference" --user-id alice

   Expected output (matches the preference that was just stored):

   [
     {
       "id": "<memory-id>",
       "memory": "Drinks a cup of black coffee every morning",
       "score": 0.62,
       "scope": "long-term",
       "categories": ["preferences"]
     }
   ]

4. List all memories for the current user.

   openclaw mem0 list --user-id alice

Step 4: Verify end-to-end read/write through TUI

openclaw tui provides a local terminal conversation interface for agents. This step verifies that memories generated through TUI conversations can be retrieved by using CLI commands, confirming the end-to-end read/write loop.

Prerequisites

• A default agent model is configured for OpenClaw, such as the Alibaba Cloud Model Studio qwen-plus series. If this is not configured, run the following command:

  openclaw onboard --auth-choice qwen-standard-api-key-cn \
                   --modelstudio-standard-api-key-cn <your-dashscope-api-key>

  Or see the OpenClaw model configuration documentation.

• The gateway is started.

  openclaw gateway restart

Verification process

1. Open TUI and have a conversation with the agent.

   openclaw tui

   Enter a preference and then exit TUI (Ctrl+C or /exit):

   User: I am left-handed.
   
   Agent: OK, I will remember that.

2. Wait for the asynchronous processing to complete, which typically takes a few seconds, and then run the following CLI commands.

   # List all memories in the current namespace
   openclaw mem0 list
   
   # Perform a semantic search with a query that does not lexically overlap with the original input
   openclaw mem0 search "hand preference"

   Expected result: The list response contains a fact extracted by the long-term memory service from the conversation. The search command uses a query that does not lexically overlap with the original input and matches the stored memory. This verifies that vectorized semantic retrieval works as expected.

The agent automatically invokes the Memory plugin during each conversation turn: before responding, it retrieves relevant memories and injects them into the context (auto-recall); after responding, it extracts new facts and writes them to the long-term memory service (auto-capture). No explicit tool invocation is required.

Automatic extraction and recall

Once configured, the plugin automatically performs the following tasks during each conversation through TUI or any connected messaging channel:

• Extraction: After the agent_end hook fires, the plugin sends conversations containing persistable user information (dietary preferences, coding habits, project backgrounds, address changes) to the AnalyticDB for PostgreSQL long-term memory service. The server-side LLM extracts structured facts and vectorizes them into the database.

• Recall: When the before_prompt_build hook fires at the start of each turn, the plugin queries the long-term memory service with the current user input. Matched memories are injected into the agent's system prompt, giving the agent cross-session memory so users do not need to repeat background information.

Search memories

openclaw mem0 search "Hangzhou"

This command returns memory entries that are semantically related to the search term. It is commonly used to check whether a write operation has taken effect or to debug the retrieval.

The retrieval uses vector similarity. The query term and the matched entries do not need to share any common words. For example, a query for accommodation preference can match memories such as prefers staying in B&Bs.

Agent tools

In addition to automatic hooks, the plugin registers the following tools that the agent can invoke explicitly during reasoning: