Agent Memory

Add persistent memory to your AI agents with remember, recall, correct, and thread turns.

Install the SDK

Memory is part of the TypeGraph SDK. Self-hosted memory uses the same adapter, embedder, and LLM setup as graph extraction; no public memory bridge constructor is required.

terminal

npm install @typegraph-ai/sdk

Initialize TypeGraph

Cloud initialization binds the client to your customer tenant. Self-hosted initialization wires the vector store, ingest embedder, optional search embedder, and LLM.

Setup example

import { TenantId, typegraphInit } from '@typegraph-ai/sdk'

const typegraph = await typegraphInit({
  apiKey: process.env.TYPEGRAPH_API_KEY!,
  tenantId: TenantId('tenant_acme'),
})

Remember

memory.remember() stores durable private memory. Put actor identity under context; TypeGraph derives a private memory graph that extends public by default.

Remember example

import { UserId } from '@typegraph-ai/sdk'

await typegraph.memory.remember(
  'Dana prefers concise answers with implementation examples.',
  {
    context: {
      userId: UserId('dana'),
    },
    graphExtraction: true,
  },
)

Recall

memory.recall() searches the caller’s private memory graph and inherited public graph context.

Recall example

const memories = await typegraph.memory.recall('How should I answer Dana?', {
  context: {
    userId: UserId('dana'),
  },
  limit: 10,
})

for (const memory of memories) {
  console.log(memory.description ?? memory.content)
}

Correct

memory.correct() applies natural-language updates while preserving provenance. Enable graphExtraction when corrections should affect the derived private graph as well as memory state.

Correct example

const correction = await typegraph.memory.correct(
  'Dana now prefers terse answers unless she explicitly asks for detail.',
  {
    context: { userId: UserId('dana') },
    graphExtraction: true,
  },
)

console.log(correction)

Track Thread Turns

Threads are ordered containers. A turn has the standard { role, content, timestamp, metadata } shape and is internally stored as a linked event for provenance and extraction.

Thread turn example

const thread = await typegraph.thread.upsert(
  {
    id: 'support-thread-123',
    name: 'Dana support session',
    description: 'Support thread about TypeGraph setup.',
    metadata: { channel: 'chat' },
  },
  { context: { userId: UserId('dana') } },
)

await typegraph.thread.addTurn(
  thread.id,
  {
    role: 'user',
    content: 'Can you help me set up TypeGraph with Next.js?',
    timestamp: new Date(),
    metadata: { source: 'web-chat' },
  },
  {
    context: { userId: UserId('dana'), threadId: thread.id },
    graphExtraction: true,
  },
)

Search Memory With Documents

search() can retrieve memory, documents, threads, events, entities, and facts together. Use resources to choose result families and weights to tune scoring.

Memory search example

const response = await typegraph.search('Help me with my TypeScript project', {
  resources: ['documents', 'threads', 'facts', 'entities'],
  weights: { semantic: 1, bm25: 0.5, graph: 0.3, recency: 0.5 },
  context: { userId: UserId('dana') },
  promptBuilder: {
    format: 'xml',
    sections: ['chunks', 'facts', 'entities'],
  },
})

const memories = await typegraph.memory.recall('Help me with my TypeScript project', {
  context: { userId: UserId('dana') },
  limit: 5,
})

console.log(response.prompt)

Back to All Guides

import { UserId } from '@typegraph-ai/sdk' await typegraph.memory.remember( 'Dana prefers concise answers with implementation examples.', { context: { userId: UserId('dana'), }, graphExtraction: true, }, )

const memories = await typegraph.memory.recall('How should I answer Dana?', { context: { userId: UserId('dana'), }, limit: 10, }) for (const memory of memories) { console.log(memory.description ?? memory.content) }

const correction = await typegraph.memory.correct( 'Dana now prefers terse answers unless she explicitly asks for detail.', { context: { userId: UserId('dana') }, graphExtraction: true, }, ) console.log(correction)

const thread = await typegraph.thread.upsert( { id: 'support-thread-123', name: 'Dana support session', description: 'Support thread about TypeGraph setup.', metadata: { channel: 'chat' }, }, { context: { userId: UserId('dana') } }, ) await typegraph.thread.addTurn( thread.id, { role: 'user', content: 'Can you help me set up TypeGraph with Next.js?', timestamp: new Date(), metadata: { source: 'web-chat' }, }, { context: { userId: UserId('dana'), threadId: thread.id }, graphExtraction: true, }, )

const response = await typegraph.search('Help me with my TypeScript project', { resources: ['documents', 'threads', 'facts', 'entities'], weights: { semantic: 1, bm25: 0.5, graph: 0.3, recency: 0.5 }, context: { userId: UserId('dana') }, promptBuilder: { format: 'xml', sections: ['chunks', 'facts', 'entities'], }, }) const memories = await typegraph.memory.recall('Help me with my TypeScript project', { context: { userId: UserId('dana') }, limit: 5, }) console.log(response.prompt)