Adapters Overview

The adapter pattern

Cognitive Memory separates the memory engine (decay, scoring, associations) from the storage backend. You pick an adapter that matches your infrastructure, and the engine works identically regardless of what’s underneath.

All adapters implement the same abstract interface (MemoryAdapter in Python, MemoryAdapter in TypeScript), providing:

• CRUD operations on memories
• Vector similarity search
• Tiered storage migration (hot/cold/stub)
• Association link management
• Consolidation helpers (find fading, mark superseded)
• Batch operations and transactions

Available adapters

Adapter	Language	Persistence	Vector search	Best for
InMemory	Python, TS	No	Brute-force cosine	Testing, prototyping
SQLite	Python	File	Brute-force cosine	Local development
Postgres	Python, TS	Yes	pgvector	Production
Redis	Python	Optional	RediSearch	Fast ephemeral storage
Convex	TypeScript	Yes	Convex vector search	Serverless apps
JSONL	TypeScript	File	Brute-force cosine	Debugging, audit logs

Decision matrix

Choosing by use case

Testing and prototyping: Use InMemoryAdapter. Zero config, instant startup, no dependencies. Data is lost when the process exits, which is fine for tests.

Local development: Use SQLiteAdapter (Python) or JsonlAdapter (TypeScript). Both persist to a local file. SQLite gives you SQL queries for debugging; JSONL gives you a human-readable append-only log.

Production single-tenant: Use PostgresAdapter with pgvector. Battle-tested, scalable, supports concurrent access. One Postgres instance can handle millions of memories with sub-100ms vector search via HNSW indexes.

Production multi-tenant (serverless): Use ConvexAdapter (TypeScript). Convex handles scaling, real-time sync, and multi-tenancy out of the box. No infrastructure to manage.

Ephemeral / cache layer: Use RedisAdapter. Sub-millisecond reads, optional persistence with RDB/AOF. Good for high-throughput scenarios where you can tolerate data loss.

Choosing by priority

Priority	Recommended adapter
Simplest setup	InMemory
Local persistence	SQLite / JSONL
Production reliability	Postgres
Lowest latency	Redis
Zero infra management	Convex
Audit trail	JSONL
Largest scale	Postgres with pgvector

Usage pattern

Python

from cognitive_memory import CognitiveMemory
from cognitive_memory.adapters.memory import InMemoryAdapter

mem = CognitiveMemory(adapter=InMemoryAdapter())

TypeScript

import { CognitiveMemory, InMemoryAdapter } from "cognitive-memory";

const mem = new CognitiveMemory({
    adapter: new InMemoryAdapter(),
    embeddingProvider: myEmbedder,
    userId: "user-123",
});

Writing a custom adapter

If none of the built-in adapters fit, you can write your own by implementing the MemoryAdapter abstract class. See Custom adapters for the full guide.