eve-memory-pg

Postgres/pgvector storage adapter for eve-memory — durable cross-session memory for Vercel eve agents.

pnpm add eve-memory eve-memory-pg pg

Your database needs the pgvector extension available (Neon, Supabase, RDS, and Vercel-integrated Postgres all ship it). The adapter runs an idempotent migration at startup.

Usage

import { defineMemory } from "eve-memory";
import { gatewayEmbedder } from "eve-memory/adapters";
import { pgMemoryAdapter } from "eve-memory-pg";

export default defineMemory({
  adapter: pgMemoryAdapter({
    connectionString: process.env.DATABASE_URL!,
    dimensions: 1536, // must match your embedding model
  }),
  embedder: gatewayEmbedder({ model: "openai/text-embedding-3-small" }),
  semanticRecall: { topK: 5, scope: "resource" },
  workingMemory: { template: "- name:\n- preferences:" },
});

Bring your own client

Anything with a query(sql, params) => Promise<{ rows }> works — a pg.Pool, Neon's serverless driver, or PGlite in tests:

pgMemoryAdapter({ client: myPool, dimensions: 1536 });

With connectionString, the adapter creates a pg.Pool and closes it on memory.dispose(). With client, you own the lifecycle.

Options

| Option | Required | Description | |---|---|---| | dimensions | ✅ | Embedding dimension of the vector column (e.g. 1536 for openai/text-embedding-3-small, 128 for stubEmbedder) | | client or connectionString | ✅ one of | How to reach Postgres | | tablePrefix | | Table name prefix, default eve_memory |

Tables

<prefix>_entries (id, resource_id, thread_id, content, vector embedding, seq, created_at) and <prefix>_working (scope, resource_id, thread_id, content, updated_at). Search uses pgvector cosine distance (<=>).

License

MIT