[Feat]: Asynchronous write-back for hybrid memory sync #432

Activity

1. eahishak

   added the

   Feature

   issue type on Nov 2, 2025
2. sscargal

   added

   keep-openPrevents the auto-close task from closing this issue.Prevents the auto-close task from closing this issue.

   on Apr 22, 2026
3. sscargal commented on Apr 22, 2026

   

   sscargal

   on Apr 22, 2026

   Last edited by sscargal

   Contributor

   More actions

   MemMachine has made extensive use of async operations.

   Here's my analysis of this feature request against the current codebase relative to this issue.

   ✅ Already addressed (partially)

   Request	Where it's implemented
   Async write methods	add_episodes in memmachine.py uses asyncio.gather(*tasks) to fan out to episodic + semantic memory concurrently
   Background worker / asyncio task group	ShortTermMemoryConsolidator.summarize() is explicitly non-blocking — it buffers episodes and runs a background asyncio.create_task loop
   Background ingestion with retry/backoff	SemanticService._background_ingestion_task() runs as a background loop with exponential backoff up to 60s
   Async R/W locking	AsyncRWLock in rw_locks.py provides full writer-priority async locking with context managers

   ⚠️ Not yet fully addressed

   The core concern — write operations blocking the API response until durable persistence completes — is still present. In memmachine.py::add_episodes:

   await asyncio.gather(*tasks)
   return episode_ids

   Writes to episodic and semantic storage are concurrent with each other (via asyncio.gather), but the API handler still awaits full completion before returning. The issue's request for a "background write queue" at the ingestion layer — where the API acknowledges receipt and returns immediately while writes are drained asynchronously — is not implemented in the hot path.

   The ShortTermMemoryConsolidator achieves this pattern for summarization, and SemanticService does it for background ingestion polling, but neither covers the primary durable-store write path.

   Conclusion

   The existing async infrastructure (locks, task groups, background workers) is excellent groundwork. What remains is a targeted enhancement: introducing a write-back queue in the add_episodes hot path so the REST handler can return episode IDs immediately and defer persistence. The issue's suggested ~25% latency improvement is a reasonable expectation given the background patterns already proven in ShortTermMemoryConsolidator.
4. sscargal

   added

   performanceIssues relating to MemMachine performanceIssues relating to MemMachine performance

   on Apr 22, 2026