What is MemoryWell?
A memory infrastructure layer for AI applications. One-page briefing — everything you need to decide whether this is for you.
What It Is
MemoryWell is a backend service that gives AI applications a place to store, retrieve, and semantically search conversations and knowledge — across sessions, across users, across time.
You call four HTTP functions. MemoryWell handles embedding, cataloguing, tier management, tenant isolation, deduplication, rate limiting, audit logging, and async queue reliability. You don't configure any of it.
mwStoreMemory after each conversation. Before the next session, it calls mwSearchMemories to find what's relevant. That's it. Your AI now has memory.What It Guarantees
Tenant Isolation
Your API key establishes your identity. You can only read and write your own memories. Cross-tenant access returns 403, not 404.
Exactly-Once Writes
Idempotency keys prevent duplicate bottles even under retry storms or duplicate enqueue events.
Audit Completeness
Every read and write is logged. Full action, timestamp, tenant_id, outcome. Nothing goes unlogged.
Recoverability
Export + import proven via DR drill. Full tenant snapshot. Hash-verified integrity. Safe to re-run.
What It Exposes — 4 Endpoints
mwStoreMemory
Store a memory. Async-queued, deduplicated, embedded, catalogued.
mwRecallMemory
Retrieve a specific memory by ID. Full conversation content included.
mwSearchMemories
Semantic search. Natural language query against all your memories.
mwListMemories
List memories by tier and time bucket. No full content returned — fast.
All endpoints: POST with JSON body. Bearer token auth. Same pattern across all four.
What It's Proven With
1000 ops soak, worker restart, Redis flush, 10× burst spike, duplicate storm, concurrent writes. All PASS.
Cross-tenant isolation, capability gates, auth invalidation, export integrity, restore, right-to-delete. All PASS.
Cohort spike test. Zero cross-tenant bleed. Zero job corruption. Production-ready verdict.
Current Known Constraints
Platform throughput ceiling
Base44 function layer limits burst writes. Async queue pacing is the mitigation — works, but rate limits apply (30 stores/min per tenant).
Search latency
Semantic search p95 can be 500–1500ms. Embedding is async post-store — first search on a new memory may miss it briefly.
Job processing is async
mwStoreMemory returns immediately with a job_id. The memory isn't queryable until the worker drains the queue (~seconds under normal load).
Beta memory cap
1,000 memories per beta tenant. 5MB max content per memory. Both are soft limits — contact for increase.
Ready to integrate?
The BetaKit has everything you need — API key, curl examples, golden path code, error reference.
Open BetaKit → Getting Started