Membase Plugin for OpenClaw

Persistent long-term memory for OpenClaw — hybrid vector search + knowledge graph.

Give your OpenClaw agent persistent memory that survives across sessions. Membase uses hybrid vector search + knowledge graph to remember not just text, but entities, relationships, and facts.

Free to start — Sign up at app.membase.so and connect in under a minute.

Install

openclaw plugins install @membase/openclaw-membase

Restart OpenClaw after installing.

Setup

openclaw membase login

Opens a browser for OAuth authentication. Tokens are saved automatically — no API keys to copy-paste. That's it, memory works automatically from here.

How It Works

Once installed, the plugin runs two hooks behind the scenes:

User message
    │
    ▼
┌─────────────────────────┐
│  Auto-Recall            │  Searches Membase for relevant memories
│  (before_agent_start)   │  and injects them as context
└───────────┬─────────────┘
            ▼
┌─────────────────────────┐
│  AI Response            │  Agent can also call membase_search,
│                         │  membase_store, etc. autonomously
└───────────┬─────────────┘
            ▼
┌─────────────────────────┐
│  Auto-Capture           │  Buffers messages, flushes to Membase
│  (agent_end)            │  for entity/relationship extraction
└───────────┬─────────────┘
            ▼
┌─────────────────────────┐
│  Membase Backend        │  Hybrid vector search + knowledge graph
│  (api.membase.so)       │
└─────────────────────────┘

Auto-Recall — Before every AI turn, searches your memories by semantic similarity and injects relevant context. Skips casual chat and short messages. Respects a maxRecallChars budget (default 4000) to avoid oversized context.
Auto-Capture — After conversations, buffers messages and sends them to Membase for extraction. Entities and relationships are automatically extracted into a knowledge graph. Flushes after 5 minutes of silence or 20 messages.
Knowledge Graph — Unlike simple vector-only memory, Membase uses hybrid vector search + knowledge graph to store entities, relationships, and facts. Search results include related nodes and edges for richer context.

AI Tools

The agent uses these tools autonomously during conversations:

Tool	Description
`membase_search`	Search memories by semantic similarity. Supports date filtering (`date_from`, `date_to`, `timezone`) and source filtering (`sources` — e.g. `['slack', 'gmail']`). Returns episode bundles with related facts and relevance scores.
`membase_store`	Save important information to long-term memory. Proactively stores preferences, goals, and context.
`membase_forget`	Delete a memory. Shows matches first, then deletes after user confirmation (two-step).
`membase_profile`	Retrieve user profile and related memories for session context.

CLI Commands

openclaw membase login              # OAuth login (PKCE) — opens browser
openclaw membase logout             # Remove stored tokens
openclaw membase search <query>     # Search memories
openclaw membase search <query> -s slack,gmail  # Filter by source
openclaw membase status             # Check API connectivity

Configuration

All configuration is managed through OpenClaw's plugin settings or ~/.openclaw/openclaw.json:

Key	Type	Default	Description
`apiUrl`	string	`https://api.membase.so`	Membase API URL. Override for self-hosted.
`tokenFile`	string	`~/.openclaw/credentials/openclaw-membase.json`	OAuth token cache file path. Stored outside the plugin directory so it survives updates.
`autoRecall`	boolean	`true`	Inject relevant memories before every AI turn.
`autoCapture`	boolean	`true`	Automatically store conversations to memory.
`maxRecallChars`	number	`4000`	Max characters of memory context per turn (500–16000).
`debug`	boolean	`false`	Enable verbose debug logs.

OAuth login keeps stable plugin config in ~/.openclaw/openclaw.json and stores rotating tokens in tokenFile.
Legacy keys (accessToken, refreshToken) are migrated automatically when present.

Enabling AI Tools

The plugin automatically adds itself to tools.alsoAllow on first load. If it doesn't take effect, restart the gateway once.

If you prefer to configure it manually, use tools.alsoAllow (not tools.allow) to avoid breaking your existing profile allowlist:

{
  "tools": {
    "profile": "coding",
    "alsoAllow": ["openclaw-membase"]
  },
  "plugins": {
    "entries": {
      "openclaw-membase": {
        "enabled": true,
        "config": {
          "autoRecall": true,
          "autoCapture": true,
          "maxRecallChars": 4000
        }
      }
    }
  }
}

"openclaw-membase" in tools.alsoAllow expands to all tools registered by this plugin and is appended on top of the active profile. Using tools.allow instead can silently break your profile allowlist if the plugin ID is not yet recognized at parse time. Without this entry, the AI still receives memory context via auto-recall but cannot call the tools explicitly.

How Membase Differs

	Simple vector memory	Membase
Storage	Flat embeddings	Hybrid: vector embeddings + knowledge graph
Search	Vector similarity only	Vector + graph traversal (entities, relationships, facts)
Extraction	Store raw text	AI-powered entity/relationship extraction
Auth	API key	OAuth 2.0 with PKCE (no secrets to manage)
Ingest	Synchronous	Async pipeline (~100ms response, background graph sync)

Development

git clone https://github.com/aristoapp/openclaw-membase.git
cd openclaw-membase
bun install
bun run check-types
bun run lint
bun run build

Contributing

Contributions welcome — see CONTRIBUTING.md for guidelines.

License

MIT

openclaw-membase