add_episodes writes the episode store before, not alongside, the episodic and semantic writes #1634

What happened

MemMachine.add_episodes (packages/server/src/memmachine_server/main/memmachine.py:735, as of 98f3cde speedkick) runs in two serial stages:

1. await episode_storage.add_episodes(...) (lines 754-759): one INSERT into the episode store, awaited alone.
2. Only after that returns: the episodic write (episodic_session.add_memory_episodes) and the semantic write (semantic_session_manager.add_message) run concurrently via asyncio.gather (lines 761-788).

Each add request therefore pays the episode-store round trip before any memory write starts. Episodic and semantic are already concurrent with each other; the episode store is not concurrent with them.

Why it cannot just be added to the gather

Both later stages consume the stored Episode, specifically its uid, and that uid is generated by the database. SqlAlchemyEpisodeStore.add_episodes inserts with insert(Episode).returning(Episode) and builds the returned models from the persisted rows (common/episode_store/episode_sqlalchemy_store.py:228-236). Neither the episodic write nor SemanticSessionManager._add_single_episode (which records episode.uid via add_message_to_sets, semantic_memory/semantic_session_manager.py:129) can start until the id exists.

Options

• Assign episode ids in the application (UUIDv7 or similar) when building the Episode from the EpisodeEntry. The three writes can then run in one gather/TaskGroup. This changes Episode.id from an autoincrement integer (get_episode, get_episodes and delete_episodes all parse ids with int()) and needs a migration plus a decision about existing integer ids.
• Keep the order but decide what happens on partial failure. Today a failure in the second stage leaves the episode row committed: reproduced with types: ["semantic"] while semantic memory is disabled, where the request returns 500 but the episode is stored. If the writes become concurrent, the same question applies in more combinations, so the failure contract should be settled together with the change.

No latency measured yet. The cost is one episode-store INSERT round trip per add request, on the critical path.

Related

• [Feat]: Asynchronous write-back for hybrid memory sync #432 asks for asynchronous write-back in general; this issue is the concrete serialization in the current add path.
• semantic_memory.enabled is honored only at startup: search, add and list still build and query the semantic stack whenever a request names the semantic type #1600: the semantic branch here also runs when semantic memory is disabled.

Investigated and written by Claude (Claude Code), filed from the account of the user who commissioned the investigation.