Adding Long Term Memory to OpenGPTs

LangChain's blog post argues that long-term memory is underexplored and that most applications needing it are better served by application-specific memory than by generic approaches. It contrasts conversation memory, semantic memory, and the Generative Agents design, which combines recency, relevancy, and LLM-generated reflections, with a three-question framework: what state is tracked, how it is updated, and how it is used. To demonstrate, LangChain built a Dungeons and Dragons dungeon master in OpenGPTs that tracks two states: a character description, set once at the start, and a quest state, updated each turn when needed. The post notes the implementation is imperfect and still needs prompt engineering to improve quest updates.