RECEIPTS PACK — STEP 3

What is MemoryWell?

Executive snapshot · 1-page · MW v1.0

What it is

MemoryWell is a persistent memory infrastructure layer for AI applications. It gives any AI platform or agent the ability to store, search, and recall memories across sessions — with full tenant isolation, correctness guarantees, and an auditable trail.

Think of it as what SMTP is to email: not the app, but the protocol. Any AI system can plug in. The memories persist. The identity continues. The trust holds.

What it guarantees

Tenant Isolation

Every byte is scoped to a single tenant. Cross-tenant reads/writes: impossible at infrastructure layer.

Correctness Invariants

Exactly-once writes via single-writer lease election. Idempotency keys prevent duplicates even under concurrent storms.

Auditability

Every operation logged to AuditLog with tenant_id, timestamp, and action type. Tamper-evident SHA-256 content hash on every bottle.

Recoverability

Export/import proven. Outbox + retry ensures zero silent drops. Repair sweep handles orphans.

What it exposes (6 key endpoints)

POST/mwStoreMemoryStore a memory (async, idempotent)

POST/mwSearchMemoriesSemantic search across tenant memories

POST/mwRecallMemoryRetrieve a specific memory by ID or key

POST/mwListMemoriesList recent memories for a tenant

POST/mwEnqueueMemoryEnqueue memory op for async processing

POST/mwGetJobStatusPoll status of an async job

What it's proven with

Production stages

Tests L–W: all pass

Cross-tenant leaks

Entropy events

Full receipts: MW Audit Ledger (Stage 11 + Stage 12) · Stage10ObservabilityDashboard

Current known constraints

·Base44 platform throughput ceiling: ~3 entity writes/sec (mitigated by async queue design)
·Semantic search latency: ~500ms–1.5s depending on embedding model round-trip
·No real-time websocket push — polling required for async job completion
·Queue workers are invoked via scheduled automation (5 min interval) — not a dedicated daemon