Integration Guide

abmind integrates with any AI tool that supports hooks, MCP, or direct Node.js embedding.

Integration Paths

Path	Best for	Setup
install-host	Kiro CLI, Claude Code, Gemini CLI, Codex CLI	abmind install-host <host>
Hermes plugin	Hermes-Agent	Copy plugin to ~/.hermes/plugins/abmind/
OpenClaw plugin	OpenClaw gateway	Install as ContextEngine plugin
MCP server	Cursor, Windsurf, Continue, Zed, any MCP client	abmind mcp (stdio)
Hooks (standalone)	Custom CLI tools, scripts	Call abmind hook-* commands
Node.js library	Embedding in your own app	import { MemoryManager } from "abmind"
abTARS (in-process)	Autonomous bridge	Built-in, no setup needed

---

Host Integration (install-host)

One command sets up lifecycle hooks, MCP server registration, and context files:

abmind install-host kiro      # Kiro CLI
abmind install-host claude    # Claude Code
abmind install-host gemini    # Gemini CLI
abmind install-host codex     # OpenAI Codex CLI

What it configures per host:

Host	Hooks	MCP	Context file
Kiro CLI	SessionStart, UserPrompt, Stop	✅ registered	.kiro/steering/
Claude Code	SessionStart, UserPromptSubmit, Stop	✅ via claude mcp add	CLAUDE.md
Gemini CLI	SessionStart, BeforeAgent, AfterAgent	✅ via extension	GEMINI.md
Codex CLI	SessionStart, UserPrompt, Stop	✅ registered	CODEX.md

Safe to re-run (idempotent). Backs up existing config before modifying. Uninstall with --uninstall.

---

Hermes-Agent Plugin

abmind implements Hermes's MemoryProvider interface — automatic recall and store on every turn without the model needing to call a tool.

# 1. Install abmind (npm or from source)
npm install -g abmind && abmind install

# 2. Copy plugin files to Hermes
mkdir -p ~/.hermes/plugins/abmind
cp <abmind-repo>/hermes-plugin/__init__.py ~/.hermes/plugins/abmind/
cp <abmind-repo>/hermes-plugin/plugin.yaml ~/.hermes/plugins/abmind/

# 3. Configure ~/.hermes/config.yaml:
memory:
  provider: abmind

# 4. Set env vars in your hermes .env:
ABMIND_HOME=/path/to/.abmind
TELEGRAM_HOME_CHANNEL=<your-chat-id>

What it provides:

Feature	How
Auto-recall	Relevant memories injected before every turn (full 7-layer pipeline)
Auto-store	Every turn recorded in background
Pre-compress capture	Saves messages before Hermes discards them
Tools	abmind_recall + abmind_store for explicit use
Sleep	Auto-registered cron (gateway mode) or manual

Unlike MCP-only integrations, the model doesn't need to decide to "save" or "recall" — it happens automatically on every turn.

Requirements: Node.js 22+, abmind on PATH, ABMIND_HOME and TELEGRAM_HOME_CHANNEL env vars set.

Known limitation: Hermes does not call prefetch() on the first turn of a new session. First turn gets wake-up context only (~200 chars). Subsequent turns get full 7-layer recall.

Architecture: The plugin is a thin routing layer — all logic lives in abmind core. Tool-originated stores default to confidence=1 (unverified). Sleep promotes to confidence=3 when corroborated.

---

OpenClaw Plugin

abmind registers as both a ContextEngine (drop-in replacement for lossless-claw) and a memory-capability plugin.

# Install in OpenClaw extensions directory
mkdir -p ~/.openclaw/extensions/abmind
cd ~/.openclaw/extensions/abmind
npm init -y && npm install abmind
echo 'export { register } from "abmind/openclaw-plugin";' > index.js

Configure in ~/.openclaw/openclaw.json:

{
  "plugins": {
    "allow": ["abmind"],
    "slots": { "memory": "abmind" },
    "entries": {
      "abmind": {
        "config": {
          "autoRecall": true,
          "autoCapture": true,
          "autoRecallMaxResults": 3,
          "sleepEnabled": true
        },
        "hooks": { "allowConversationAccess": true }
      }
    }
  },
  "tools": { "allow": ["abmind_recall", "abmind_store"] }
}

What registers:

• ContextEngine: ingest, assemble, compact, afterTurn
• Memory-capability: promptBuilder, runtime (MemorySearchManager), publicArtifacts
• Agent tools: abmind_recall + abmind_store
• Lifecycle hooks: autoRecall (before_agent_start), autoCapture (agent_end)

---

MCP Server

abmind mcp    # starts on stdio — connect from any MCP client

Works with any tool that supports the Model Context Protocol: Cursor, Windsurf, Continue, Zed, VS Code extensions, custom clients.

Available tools

Tool	Description
memory_recall	Search memories by keywords/semantic query
memory_store	Store a new memory (fact, preference, entity)
memory_edit	Edit or delete an existing memory
memory_status	Memory system stats (counts, DB size)
memory_wakeup	Get current wake-up context

Registering manually

If your tool doesn't support install-host, register the MCP server directly:

{
  "mcpServers": {
    "abmind": {
      "command": "abmind",
      "args": ["mcp"]
    }
  }
}

---

Hook Commands (Standalone)

For custom integrations, call these from your tool's hook/event system:

Hook	When to call	What it does
abmind hook-wakeup	Session start	Outputs wake-up context (recent facts, profile)
abmind hook-recall	Before each prompt	Outputs relevant memories for the current query
abmind hook-store	After each response	Records the user+assistant turn
abmind hook-doctor	On demand	Diagnoses hook config issues

Each outputs text to stdout — inject into the model's system prompt or context.

---

Node.js Library (Embedding)

For building your own memory-powered application:

import { createMemoryBackend } from "abmind";
import { loadMemoryConfig } from "abmind";

const config = loadMemoryConfig();
const backend = await createMemoryBackend(config);

// Store
backend.recordMessage({ role: "user", content: "I live in Budapest", userId: "user-1", sessionId: "s1", timestamp: Date.now() });

// Recall
const results = await backend.recall({ translated: ["Budapest", "location"], userId: "user-1", limit: 5 });

backend.close();

---

abTARS (In-Process)

When running as part of abTARS, abmind is loaded in-process via file:../abmind dependency. No separate setup — abtars install handles everything. Memory is available to all agents (professor, dreamy, browsie, coding) and the heartbeat/sleep system.