gharibyan/agent-memory

This is a TypeScript library called agent-memory-sdk, available as a single npm package, that adds persistent memory to AI agent applications. When you build a chatbot or AI assistant, one common problem is that the AI has no way to remember past conversations unless you manually paste the entire chat history into every new request. This library wraps your model calls so that memory is stored, recalled, and updated automatically, without you having to manage that history yourself. The memory system is organized into layers. User memory is tied to a specific person and persists across all their sessions. Thread memory covers a single conversation. Operation memory holds compact context for a specific task or workflow. You can set a budget to limit how much stored context gets injected into any one call, which keeps token costs under control. You can also turn off recall or learning for individual calls when you need a clean slate. For storing the memories, the library defaults to a local JSON file on disk, which is fine for prototypes. It also ships adapters for SQLite (a lightweight local database) and for Postgres with pgvector, which is a Postgres extension that enables vector search. The Postgres adapter handles its own database migrations automatically on first use, including creating the search index and relational tables. The library works with several AI model providers. Official adapters are included for OpenAI, Anthropic, Gemini, and xAI. There is also a generic helper for any provider that exposes an OpenAI-compatible chat API, so you can point it at custom or third-party model endpoints by supplying a base URL and API key. Internally the repository is split into multiple packages covering the core engine, each storage adapter, and each provider adapter, but all of those are compiled and bundled together into the single public npm package. End users install one package and get everything. The project is open source under the MIT license.