Skip to main content
The MemoryOS quality gate runs before memory ingestion is queued. Its job is to prevent low-value, duplicate, or out-of-budget writes from becoming stored memories.

​
The four layers

LayerPurposeBackend rule
L1Per-user rate limitingBlocks when requests exceed the tenant’s per-user per-minute limit
L2Low-quality input filterBlocks when the quality score is below 0.35
L3Semantic duplicate detectionBlocks when semantic similarity is above 0.92
L4Budget governanceBlocks when quota is exhausted and overage policy is block

​
What gets blocked

LayerExample inputTypical reason
L1Same user spams many writes in one minuterate_limit_exceeded
L2"ok", "hi", "??"low_quality
L3Re-sending the same preference statement repeatedlyduplicate_query
L4Tenant is out of monthly calls or tokens on block modebudget_exhausted

​
How quality is scored

The current quality score is based on:
  • number of messages
  • average message length
  • lexical diversity
  • whether the conversation contains a question signal
Very short or content-free messages are likely to score poorly.

​
How to improve your block rate

  1. Send meaningful user facts, preferences, goals, or procedures instead of filler text.
  2. Avoid writing the same memory-worthy statement repeatedly.
  3. Batch coherent conversational turns together instead of sending single-word fragments.
  4. Respect the per-user write rate limit.
  5. Monitor blocked_reason and budget_remaining_pct on add() responses.

​
Operational note

A blocked write returns HTTP 200, but the response status indicates the blocking layer, such as L2 or L4. Inspect the response body, not just the status code.