Thing Event SystembyPentatonic
OpenClaw pluginMITPart of the AI Agent SDK

Context engine for
OpenClaw agents.

Slot the Pentatonic memory plugin into your OpenClaw stack and every session gets persistent, searchable memory. Retrieval fires before each model run; decay and consolidation run on the right cadence; storage is local Postgres or hosted TES — your choice.

View on GitHubBack to SDK

Install

One command. Conversational setup.

1. Install the plugin

openclaw plugins install @pentatonic-ai/ai-agent-sdk

Pulls the plugin into your OpenClaw environment. The package ships both the Claude Code hooks and the OpenClaw context engine; OpenClaw picks up the right one.

2. Tell OpenClaw to set it up

Set up pentatonic memory

The agent will ask whether you want local (Docker — fully private, runs on your machine) or hosted (Pentatonic TES — team-shared memory + analytics). For hosted mode it walks you through account creation, email verification, and API key setup.

Or use the CLI directly: openclaw pentatonic-memory local.

Lifecycle

Four hooks, every turn

Unlike MCP tools (agent-driven, optional), the context engine fires deterministically on every lifecycle event. Memory just happens.

ingest

Fires when a message is added to the session. Stores user/assistant content as a memory with embedding + HyDE query expansion.

assemble

Fires before each model run. Searches for relevant memories using the latest user message and returns them as a system-prompt addition.

compact

Fires when the context window fills. Runs the decay cycle to evict low-confidence memories.

afterTurn

Fires after each turn completes. Runs the consolidation pass — frequent memories get promoted from episodic to semantic.

Configuration

Local or hosted — same plugin

Drop one of these blocks into your openclaw.json. The conversational setup writes it for you.

Local memory

openclaw.json
{
  "plugins": {
    "slots": { "contextEngine": "pentatonic-memory" },
    "entries": {
      "pentatonic-memory": {
        "enabled": true,
        "config": {
          "database_url": "postgres://memory:memory@localhost:5433/memory",
          "embedding_url": "http://localhost:11435/v1",
          "embedding_model": "nomic-embed-text",
          "llm_url": "http://localhost:11435/v1",
          "llm_model": "llama3.2:3b",
          "client_id": "default",
          "search_limit": 5,
          "min_score": 0.3
        }
      }
    }
  }
}

Run npx @pentatonic-ai/ai-agent-sdk memory first to spin up the Docker stack (Postgres + pgvector + Ollama).

Hosted TES

openclaw.json
{
  "plugins": {
    "slots": { "contextEngine": "pentatonic-memory" },
    "entries": {
      "pentatonic-memory": {
        "enabled": true,
        "config": {
          "tes_endpoint": "https://your-client.api.pentatonic.com",
          "tes_client_id": "your-client",
          "tes_api_key": "tes_your-client_xxxxx"
        }
      }
    }
  }
}

Get credentials at /signup or via npx @pentatonic-ai/ai-agent-sdk init.

Tunables

Sensible defaults; override when you need to

client_iddefault

Memory namespace. Shared across all sessions on this OpenClaw instance.

search_limit5

How many memories to inject per prompt. Higher = more context, but more tokens used.

min_score0.3

Similarity floor for retrieval. Below this, memories are filtered out as not-relevant-enough.

embedding_modelnomic-embed-text

Local Ollama embedding model. 768-dim by default. Use mxbai-embed-large for 1024-dim if you want more semantic surface area.

llm_modelllama3.2:3b

Local chat model used for HyDE query expansion. Pi 5 with 8GB RAM runs this comfortably.

OpenClaw plugin

Drop-in memory for your OpenClaw agents

Free for local mode (your Docker, your data). Hosted TES adds team-shared memory and analytics on the same per-token meter as the proxy.

Get API keyBrowse the SDK