agent-memory · packages

Provides a unified interface for managing AI agent long-term memory, including automated fact extraction, semantic retrieval, and lifecycle policies. It exposes an `AgentMemory` class that integrates storage backends, embedding providers, and LLM-based extraction logic.

Provides the canonical TypeScript interfaces, enums, and utility functions for the agent-memory ecosystem, including vector similarity calculations and retry logic. It serves as the shared type definition layer for building custom memory adapters and agent integrations.

Provides a unified interface for generating text embeddings using OpenAI, Cohere, or HuggingFace providers. It includes a decorator class for transparently caching results in memory to reduce API calls and latency.

Provides an in-memory event bus and a set of TypeScript interfaces for hooking into agent memory lifecycle events like storage, retrieval, and contradiction resolution. It exposes an `InMemoryEventBus` class for single-process use and a `MemoryEventBus` interface for implementing custom pub/sub adapters.

Extracts structured facts, preferences, and decisions from conversation logs using LLMs and generates corresponding vector embeddings. It provides a `MemoryExtractor` class that requires separate LLM and embedding provider implementations to process conversation turns.

Provides a unified interface for LLM text completion and structured JSON output. It includes a pre-built class for OpenAI-compatible APIs and allows custom implementations via the `LLMProvider` interface.

Provides a `PolicyEngine` class to manage the lifecycle of agent memories through configurable rules for exponential decay, automated forgetting, and contradiction resolution. It integrates with `@reaatech/agent-memory-core` to evaluate memory state and resolve conflicts based on strategies like recency or confidence.

Provides a `MemoryRetriever` class to query and rank stored memories using semantic, temporal, and importance-based strategies, alongside a `ContextInjector` to format results for LLM prompts. It requires an external memory storage provider and an embedding provider to function.

Provides a unified interface for persisting and querying agent memories, offering both an in-memory implementation and a PostgreSQL adapter with pgvector support. It includes a fluent query builder class for filtering and performing semantic similarity searches.