mnem-langgraph

A LangGraph BaseStore backed by Mnemosyne. The agent's long-term memory gets history: every put is a commit, and you can ask why a memory says what it does, or branch to test a hunch.

from mnem_langgraph import MnemosyneStore

store = MnemosyneStore("./agent-memory")          # one .mnem store
store.put(("memories", "u1"), "plan",
          {"tier": "enterprise", "_meta": {"source": "ticket-4821"}})

item  = store.get(("memories", "u1"), "plan")     # -> Item
hits  = store.search(("memories",), query="enterprise")
blame = store.why(("memories", "u1"), "plan")     # which commit, and from where

Pass it to a graph the usual way (builder.compile(store=store)), or use it directly from a node.

Mapping

• namespace + key -> node id ":".join([*namespace, key]). A ":" inside a segment is not supported.
• value -> node content. A reserved _meta key (source / step / observation / tool_call / note) is lifted into provenance so blame works; it is stored too, so get round-trips.
• search is a plain substring (query) and equality (filter) match, not semantic. Wrap a vector store if you need ranking.
• Async (aget / aput / asearch) runs the sync path in a worker thread.

Extra methods

Not on the BaseStore interface, for a graph node to call: branch(name), switch(target), why(namespace, key), history(limit=...).

See ADR-0016.