Operations Pack
SLO targets, alert rules, on-call playbook, cost model, platform ceilings and mitigation.
SLO Targets — MW v1.0 Beta
| Endpoint | p50 target | p95 target | p99 target | Notes |
|---|---|---|---|---|
| mwStoreMemory (enqueue) | <200ms | <500ms | <2000ms | Enqueue only — worker processes async |
| mwRecallMemory | <150ms | <400ms | <1000ms | Direct ID fetch from entity layer |
| mwSearchMemories | <400ms | <1500ms | <3000ms | Embedding similarity — inherently slower |
| mwListMemories | <150ms | <400ms | <800ms | Catalog filter — no content returned |
| Queue drain (worker) | <5s | <30s | <120s | End-to-end store to queryable |
Stage 11 soak confirmed p95 <2s for store, p95 <800ms for recall and list. Search latency is higher due to embedding round-trip.
Alert Rules
On-Call Playbook
Cost Model v1 — Beta
| Operation | Cost components | Notes |
|---|---|---|
| mwStoreMemory | 1 Base44 function call + 1 OpenAI embedding + ~3 entity writes | Heaviest op. Async queue amortises cost over time. |
| mwRecallMemory | 1 function call + 1 entity read | Cheapest retrieval op. Scales cleanly. |
| mwSearchMemories | 1 function call + 1 OpenAI embedding + vector scan | Second most expensive. Limit results to ≤10 per call. |
| mwListMemories | 1 function call + 1 entity query (no embedding) | Very cheap. Good for pagination and navigation. |
Platform Ceilings + Mitigation
CEILING
30 stores/min per tenant
MITIGATION
Async queue pacing. Batch enqueues. Backpressure on client side.
CEILING
20 searches/min per tenant
MITIGATION
Cache search results client-side for repeated queries. Deduplicate before searching.
CEILING
Base44 function timeout (~30s)
MITIGATION
Worker processes jobs in small batches. No single job should approach timeout.
CEILING
5MB max content per memory
MITIGATION
Chunk large conversations before storing. Use mwChunkLargeConversation if available.
CEILING
1,000 memories per beta tenant
MITIGATION
Tier management (HOT/WARM/COLD/FROZEN). Archive old memories. Contact for limit increase.