@engram-mem/openai

Generates dense vector embeddings for text using OpenAI's embedding models (text-embedding-3-small, text-embedding-3-large). Integrates with Engram's memory system to convert unstructured text into fixed-dimensional vectors suitable for similarity search and retrieval. Handles batch processing and caches embeddings to avoid redundant API calls.

Leverages OpenAI's language models to produce summaries of long-form text in both extractive (selecting key sentences) and abstractive (generating new summary text) modes. Integrates with Engram's memory to compress conversation history and long documents into concise representations while preserving semantic meaning. Supports configurable summary length and style parameters.

Extracts structured entities (people, organizations, locations, concepts, dates) from unstructured text using OpenAI's language understanding capabilities. Automatically tags memories with extracted entities to enable entity-based retrieval and relationship mapping. Supports custom entity schemas and hierarchical entity relationships.

Applies OpenAI-powered cross-encoder models to rerank retrieved memories based on semantic relevance to a query. Unlike embedding-based similarity (which scores independently), cross-encoders jointly encode query and candidate text to produce more accurate relevance scores. Integrates with Engram's retrieval pipeline to refine initial embedding-based results before returning to the agent.

Automatically selects and prioritizes memories to include in agent context based on relevance, recency, and importance scores. Uses embeddings, entity relationships, and summarization to fit the most valuable information within token budgets. Implements a multi-level memory hierarchy (working memory, episodic memory, semantic memory) with intelligent promotion/demotion based on access patterns.

Converts raw conversation transcripts into structured memory artifacts by applying embeddings, summarization, entity extraction, and metadata enrichment in a coordinated pipeline. Handles multi-turn conversations, speaker attribution, and context preservation. Stores results in Engram's memory format with full indexing for later retrieval.

Provides abstraction layer allowing Engram to work with different embedding, summarization, and extraction providers (OpenAI, Anthropic, local models) through a unified interface. Enables switching providers without changing agent code. Handles provider-specific API differences, error handling, and fallback strategies.