Vector memory adapters

Drop-in VectorMemory for pgvector, Pinecone, Qdrant, Chroma, and Upstash Vector.

@agentskit/memory ships five new VectorMemory implementations. Each targets a different deployment story — SQL-native (pgvector), serverless HTTP (Pinecone, Upstash), self-hosted REST (Qdrant, Chroma). All obey the same three-method contract (store / search / delete), so you can A/B providers without touching agent code.

#Install

npm install @agentskit/memory

#Postgres + pgvector

BYO SQL runner so you pick the driver (pg, postgres, @neondatabase/serverless, Supabase client).

import { pgvector } from '@agentskit/memory'
import { Pool } from 'pg'

const pool = new Pool({ connectionString: process.env.DATABASE_URL })
const memory = pgvector({
  runner: {
    query: async (sql, params) => {
      const r = await pool.query(sql, params)
      return { rows: r.rows }
    },
  },
  table: 'agentskit_vectors',
})

Expects a table like:

CREATE TABLE agentskit_vectors (
  id text primary key,
  content text,
  embedding vector(1536),
  metadata jsonb
);

#Pinecone

import { pinecone } from '@agentskit/memory'

const memory = pinecone({
  apiKey: process.env.PINECONE_API_KEY!,
  indexUrl: 'https://<idx>-<project>.svc.<region>.pinecone.io',
  namespace: 'prod',
})

#Qdrant

import { qdrant } from '@agentskit/memory'

const memory = qdrant({
  url: process.env.QDRANT_URL!,
  apiKey: process.env.QDRANT_API_KEY,
  collection: 'agents',
})

#Chroma

import { chroma } from '@agentskit/memory'

const memory = chroma({
  url: 'http://localhost:8000',
  collection: 'agents',
})

#Upstash Vector

import { upstashVector } from '@agentskit/memory'

const memory = upstashVector({
  url: process.env.UPSTASH_VECTOR_URL!,
  token: process.env.UPSTASH_VECTOR_TOKEN!,
})

#Shared contract

All five implement:

interface VectorMemory {
  store(docs: VectorDocument[]): Promise<void>
  search(embedding: number[], opts?: { topK?: number; threshold?: number }): Promise<RetrievedDocument[]>
  delete?(ids: string[]): Promise<void>
}

Results include a normalized score in [0, 1] (higher is better). pgvector converts cosine distance; Chroma converts 1 - distance; Pinecone / Qdrant / Upstash pass through the native score.

#See also

RAG reranking — wrap any of these with BM25 hybrid
Hierarchical memory — use as the recall tier
Encrypted memory — layer on top for zero-trust

Explore nearby

✎ Edit this page on GitHub·Found a problem? Open an issue →·How to contribute →