Short-term memory is process-local: recent episodes are never reloaded, replicas diverge, and the persisted summary is last-writer-wins #1748

What happened

ShortTermMemory keeps its recent episodes in a deque in process memory (episodic_memory/short_term_memory/short_term_memory.py:123-129). create() reloads only the summary and carries the comment # ToDo: Retrieve the episodes from raw data storage (:157-173). The consolidator's summary, pending episodes and worker task are per instance (:502-524). The summary is persisted by save_short_term_memory, which is SELECT then UPDATE-or-INSERT with no upsert and no lock (common/session_manager/session_data_manager_sql_impl.py:378-424).

• With the default session_manager.instance_cache_size: 0 (common/configuration/__init__.py:75-79) the instance closes when its last reference drops (episodic_memory/instance_lru_cache.py:208-218), so even on one process every request starts with an empty recent window. Only the summary survives, and only if summarization ran.
• With a cache, each replica holds its own deque for the same session and answers short-term queries from the episodes it received itself. Each replica's consolidator builds a summary on top of its own copy and overwrites the others'. Two replicas' first saves can both INSERT; the loser's IntegrityError is not in the caught set (:647) and reaches the summary loop's catch-all, so that replica's summary is never set.
• ShortTermMemory.create calls data_manager.create_tables() on every instance build (:159-162), which with cache size 0 is every add or search. Schema provisioning runs from every process's boot path: create_all races on cold boot and never evolves a table, Alembic runs destructive migrations on first use #1570 covers the DDL side.

Expected

Either short-term memory is declared a single-process component and a multi-replica deployment refuses or disables it, or its state lives in the database: episodes reloaded on open, summary saved by an upsert keyed on session and sequence. #1719 covers its cancellation and reset defects; this issue is the cross-replica one.

Notes

Code read at d6068cdbf (main), paths under packages/server/src/memmachine_server/. Short-term memory is not needed for horizontal scaling, so removal is acceptable; the issue records the decision so it is explicit rather than silent.

🤖 Written by Claude Code (Claude Fable 5.1) on behalf of @edwinyyyu.