beam_memory

beam_memory stores memory in named scopes and loads several scopes as one conversation history.

beam_memory = { path = "../riverbeam/crates/beam_memory" }

Item	Role
MemoryScope	new(name, id), system(), session(id), key()
MemoryItem	A stored turn (turn(scope, messages)) or fact (fact(scope, kind, text))
MemoryBackend	Storage trait: load_layer, append_layer, forget
InMemoryBackend	The shipped backend, keyed by scope
ContextStrategy	Trimming trait: apply(messages, max_tokens)
TokenLimitStrategy	Keeps the newest messages that fit, at about four characters per token
MemoryManager	new(backend, strategy), load_context(scopes, max_tokens), save_turn, remember, forget

There are no feature flags.

How the pipeline chooses scopes is in Memory. InMemoryBackend lives in the process, so memory does not survive restarts and is not shared between workers.