Memory systems#
Most actors need to store their observations and make use of them when queried. InterLab offers a common BaseMemory interface along with several implementations:
interlab.actor.memory.ListMemorysimply keeps all observations in a list, and every query has all of them available. Of course, this may run out of LLM context space in longer scenarios. Also works for non-textual observations (e.g. in GT contexts).interlab.actor.memory.experimental.SummarizingMemorysummarizes earlier and less terse textual observations to keep all observations fit a given overall token capacity. Marked as experimental for now as it has not been thoroughly battle-tested.interlab.actor.memory.experimental.SimpleEmbeddingMemoryembeds every observation using an embedding language model. Given a query, it embeds the query and returns up to the given number or length of relevant memories (ordered temporally). Similarly not as well tested in our experiments yet.
See the source or the API docs for details of each of the memories.
Memory and actors#
ActorWithMemory is intended for agents that directly pass their observations to their memory. ActorWithMemory.__init__ takes a memory argument, which defauls to a new ListMemory. This is not meant to be a strict hierarchy - you can add memory to any of your own actor implementations, and actors may even perform more than mere recording to memory upon getting an observation.