Framework Integrations

Engram is framework-agnostic, but adapters and patterns are available for common agent stacks.

LangChain Pattern

Use Engram tool wrappers to persist and recover intermediate decisions.

import { EngramStoreTool, EngramRetrieveTool } from 'engramjs/adapters/langchain'

const tools = [new EngramStoreTool(sdk), new EngramRetrieveTool(sdk)]

Best practice:

• persist decision snapshots at stable checkpoints
• store CID references in your agent state machine

LlamaIndex Pattern

Use reader/writer abstractions for document or memory flow.

import { EngramWriter, EngramReader } from 'engramjs/adapters/llamaindex'

const writer = new EngramWriter(sdk)
const cid = await writer.store(document)

const reader = new EngramReader(sdk)
const reloaded = await reader.loadData(cid)

MCP-First Integration

When your host supports MCP natively, you can integrate without writing SDK glue code:

1. run @engram-sdk/mcp
2. register server URL in client config
3. call engram_* tools from your AI runtime

This is often the quickest path to production experiments.

Context Window Offload Pattern

A practical integration pattern from the reference agent loop is to offload context once token usage crosses a threshold:

const MAX_TOKENS = 8192
const OFFLOAD_THRESHOLD = 0.8

if (tokenCount / MAX_TOKENS >= OFFLOAD_THRESHOLD) {
  const receipt = await sdk.store(
    { messages, iteration },
    { priority: 'standard' }
  )

  messages = [{
    role: 'system',
    content: `Previous context was offloaded to CID ${receipt.cid}.`
  }]
}

Why this helps:

• prevents runaway prompt growth
• preserves recoverability with explicit CID references
• gives deterministic points for renew/prune policies later