NBJ OB1 Agent Memory for OpenClaw

Governed Nate Jones OB1 memory for OpenClaw: recall before the task, write back after, inspect everything.

Built by Nate B. Jones / OB1. Follow Nate for practical AI systems, agent workflows, and implementation notes: Substack and natebjones.com.

Install

Current OpenClaw 2026.5.2 installs the ClawHub-hosted package cleanly from the published tarball:

curl -fsSL \
  https://clawhub.ai/api/npm/@natebjones/ob1-agent-memory/-/natebjones-ob1-agent-memory-0.1.4.tgz \
  -o natebjones-ob1-agent-memory-0.1.4.tgz

openclaw plugins install ./natebjones-ob1-agent-memory-0.1.4.tgz

The package is also published on ClawHub as @natebjones/ob1-agent-memory. When OpenClaw's clawhub: resolver accepts npm-pack artifact metadata directly, this should become the normal one-line install:

openclaw plugins install clawhub:@natebjones/ob1-agent-memory

For local linked development:

npm install --ignore-scripts --omit=peer
npm run build
openclaw --profile ob1-agent-memory plugins install . --link

Required Config

{
  secrets: {
    providers: {
      ob1_agent_memory: {
        type: "file",
        path: "/path/to/ob1-agent-memory-key",
        mode: "singleValue"
      }
    }
  },
  tools: {
    allow: [
      "group:openclaw",
      "openbrain_recall",
      "openbrain_writeback",
      "openbrain_report_usage",
      "openbrain_inspect_memory",
      "openbrain_list_review_queue",
      "openbrain_review_memory",
      "openbrain_get_recall_trace"
    ]
  },
  plugins: {
    entries: {
      "nbj-ob1-agent-memory": {
        config: {
          endpoint: "https://YOUR_PROJECT_REF.supabase.co/functions/v1/agent-memory-api",
          accessKey: {
            source: "file",
            provider: "ob1_agent_memory",
            id: "value"
          },
          workspaceId: "workspace_123",
          projectId: "project_456"
        }
      }
    }
  }
}

Configure secrets.providers.ob1_agent_memory with a file or exec provider before enabling the plugin. The plugin resolves OpenClaw SecretRefs at tool execution time so the access key does not need to live in plaintext config. Env SecretRefs are intentionally not used in the launch package so OpenClaw's install-time safety scan does not classify the memory client as env-to-network credential forwarding.

Current OpenClaw builds also require explicit tools.allow entries before plugin tools are exposed to the model. plugins inspect --runtime can show the plugin is registered even when the agent cannot see the tools, so run a native tool smoke test after enabling the plugin.

Install the paired skill from ClawHub or use the bundled plugin skill so agents respect OB1 provenance, review, and use-policy rules.

Local validation target: openclaw --profile ob1-agent-memory plugins inspect nbj-ob1-agent-memory --runtime --json should list all seven openbrain_* tools and no diagnostics.

Native smoke target: run an OpenClaw agent turn that calls openbrain_list_review_queue with no shell/file tools. The result should show an openbrain_list_review_queue tool call and zero failures.

For the repeatable full loop harness:

OPENCLAW_BIN="/home/lej/.local/bin/openclaw" \
OPENCLAW_PROFILE="personal" \
OPENCLAW_AGENT="jonathan" \
npm run smoke:native

The harness enables the plugin, verifies runtime registration for all seven openbrain_* tools, runs a native agent turn that calls every OB1 tool, parses the session transcript for non-OB1 tool calls and tool errors, and disables the plugin again by default.