Storage & Memory (Persistence)

Concepts: State Management

LLMs are stateless. They don't remember you. To build a conversation, you need to manage state in two ways:

1. Memory (Chat History): Short-term context. "What did we just talk about?"
2. Storage (Key-Value): Long-term persistence. User preferences, session data, or cached results.

---

1. Memory Adapters

Conversation History

Memory adapters provide a standard interface to load() past messages and save() new ones.

When to use what?

• LangChain (BaseListChatMessageHistory): The Industry Standard. LangChain has adapters for Redis, Postgres, Mongo, DynamoDB, and dozens more. If you need to save chat logs to a real database, use a LangChain history adapter.

  • Upstream Docs: BaseListChatMessageHistory

• In-Memory (Simple): For quick scripts or stateless serverless functions, you might just pass an array of messages directly to the Workflow run() method. You don't always need an adapter.

---

2. Key-Value Stores

Arbitrary Data

Sometimes you need to save more than just messages—like a user's uploaded PDF ID, a session token, or a serialized workflow state.

• Interface: KVStore
• Why use it?: It abstracts the backend. You can write your logic against store.get(key) and swap Redis for In-Memory for S3 without changing your business logic.

Example: Redis Storage

We recommend passing LangChain Stores into our adapter.

import type { KVStore } from "@geekist/llm-core/adapters";

---

3. Caching

Performance & Cost Savings

Caching allows llm-core to remember the result of an expensive operation (like a GPT-4 generation or a long embedding) and return it instantly next time.

How it works

We utilize a simple Cache interface that can sit in front of any operation.

// #region docs
import { createMemoryCache } from "@geekist/llm-core/adapters";
import type { Cache } from "@geekist/llm-core/adapters";

// 1. Create a cache (TTL is provided at call time)
const cache: Cache = createMemoryCache();

// 2. Use it in a workflow (example conceptual usage)
// The workflow engine checks this cache before calling the model
// #endregion docs
void cache;

We also support Persistent Caching by wrapping a KVStore:

import { createCacheFromKVStore } from "@geekist/llm-core/adapters";
import type { Blob, Cache, KVStore } from "@geekist/llm-core/adapters";

---

Supported Integrations (Flex)

Capability	Ecosystem	Adapter Factory	Upstream Interface	Deep Link
Memory	AI SDK	fromAiSdkMemory	MemoryProvider	(Experimental)
Memory	LangChain	fromLangChainMemory	BaseListChatMessageHistory	Docs
Memory	LlamaIndex	fromLlamaIndexMemory	BaseMemory	Docs
KV Store	LangChain	fromLangChainStore	BaseStore	Docs
KV Store	LlamaIndex	fromLlamaIndexDocumentStore	BaseDocumentStore	Docs
Cache	AI SDK	fromAiSdkCacheStore	(Custom)	-
Cache	LangChain	fromLangChainStoreCache	BaseStore	Docs
Cache	LlamaIndex	fromLlamaIndexKVStoreCache	BaseKVStore	Docs