MemoryLabs is a single secure connector (MCP) that every major assistant plugs into. They write to and read from one memory that lives in your vault, and a frontier model quietly consolidates it every night. Here is the whole system, drawn honestly.
Ingress · one connector, every assistant
ClaudeChatGPTGeminiGrokCursor→MCP connector · OAuth 2.1resolves per-user identity, then filters every query by it
every raw memory→frontier synthesis · Sonnet→transactional apply→living profile
Store · the boundary that matters
Postgres + pgvectorper-user RLS isolationhashed tokensyou own every row
Reads never wait on a frontier model, the understanding was precomputed overnight. Writes never lose a memory, intake fails over across providers and falls back to a raw store. Every query is scoped to you at the database, not just the app.
The three loops, in words
Writesynchronous
Save without losing anything.
A memory passes a credential firewall, is split into atomic facts and classified, embedded, and filed with semantic dedupe. If one AI provider is down, another does the same work; if both fail, it's still stored raw. Nothing is ever dropped.
Readsynchronous
Recall never waits on a model.
A query runs hybrid vector plus keyword search, fused and freshness-weighted, against a profile that was already synthesized overnight. No frontier model sits in the read path, so recall is fast and predictable.
Organizenightly · batch
Your memory sleeps on it.
Each night a frontier model reads what's new and consolidates it, merging duplicates, resolving contradictions, and writing a living profile, applied to your vault in one transaction. You keep the final say and can read every version.