@lumetra/engram-agentkit

Durable, explainable memory tools for Inngest AgentKit — powered by Engram by Lumetra.

Drop these tools into any createAgent({...}) and your agent gains six memory operations: store, query (semantic + graph retrieval), list, and delete.

Install

npm install @lumetra/engram-agentkit @inngest/agent-kit zod

Get an API key at https://lumetra.io.

Quickstart

import { createAgent, anthropic } from "@inngest/agent-kit";
import { createEngramTools } from "@lumetra/engram-agentkit";

const tools = createEngramTools({
  apiKey: process.env.ENGRAM_API_KEY!,
  bucket: "my-app",                       // optional default bucket
  // baseUrl: "https://api.lumetra.io",   // override for self-hosted
});

const agent = createAgent({
  name: "Memory-aware assistant",
  system:
    "You have durable memory. Call `query_memory` before answering anything " +
    "that might depend on prior context, and call `store_memory` whenever the " +
    "user shares a fact, preference, or decision worth remembering.",
  model: anthropic({
    model: "claude-sonnet-4-5",
    defaultParameters: { max_tokens: 32768 },
  }),
  tools,
});

const { output, toolCalls } = await agent.run(
  "Remember that my favorite framework is Inngest. Then tell me what you remembered.",
  { maxIter: 1 }  // see note below
);

console.log(output);
console.log("tool calls:", toolCalls.length);

Heads-up: maxIter and the Anthropic adapter (AgentKit 0.13.2). If maxIter >= 2, AgentKit's Anthropic adapter throws tool_use blocks can only be in assistant messages mid-loop — the bug is in how it serializes the tool result message back to Claude. This is an upstream AgentKit bug, not our package. Until it's fixed, keep maxIter: 1 for Anthropic + tool flows. OpenAI / Gemini paths aren't affected. We caught this in our e2e smoke against the published artifact.

Tools

| Tool | What it does | | ----------------- | ---------------------------------------------------------------------------- | | store_memory | Save a fact / preference / decision (content, bucket?) | | query_memory | Semantic + graph retrieval over a bucket (query, bucket?) | | list_memories | Recent raw memories (bucket?, limit?) | | list_buckets | All buckets visible to the API key (limit?, offset?) | | delete_memory | Delete by ID (memory_id, bucket?) | | clear_memories | Wipe a whole bucket (bucket?) — destructive |

If you set bucket in createEngramTools({ bucket: "..." }), the agent can omit bucket on every call and the default is used. Otherwise the agent must pass it explicitly.

Configuration

createEngramTools({
  apiKey: string,         // required (or pass `client`)
  bucket?: string,        // default bucket for tools that omit one
  baseUrl?: string,       // default "https://api.lumetra.io"
  fetch?: typeof fetch,   // custom fetch implementation
  client?: EngramClient,  // advanced: bring your own client
});

You can also use the raw client directly:

import { EngramClient } from "@lumetra/engram-agentkit";

const engram = new EngramClient({ apiKey: process.env.ENGRAM_API_KEY! });
await engram.store("my-app", "Jacob prefers TypeScript over JavaScript.");
const hits = await engram.query("my-app", "What does Jacob prefer?");

Privacy

See PRIVACY.md. The adapter sends only the tool arguments to the Engram API; no other agent state or files leave your process.

License

MIT - Lumetra Labs, Inc.