Data isolation

Mneva's isolation model is structural, not procedural. There is no shared row, no shared table, no cross-tenant WHERE clause anyone could forget.

One brain per tenant

Every customer's brain is its own SQLite database file on the engine box:

/var/lib/mneva/tenants/
  registry.json            (id, name, key, created_at — no contents)
  t_3194131b2393.db        (tenant A's brain — memories, beliefs, predictions, dreams, evidence)
  t_3194131b2393.db-shm    (WAL shared memory)
  t_3194131b2393.db-wal    (WAL log)
  t_80744acd0d9b.db        (tenant B — entirely separate file)
  ...

When a request comes in with x-mneva-key: mnv_..., the engine looks up the tenant in the registry, resolves it to a database file path, and opens that file. There is no cross-tenant query path.

This is harder to get wrong than a shared table with tenant_id columns. A bug in a query cannot leak data between tenants because the query is running against the wrong physical database — and would have to be redirected at file-system level, not at SQL level.

What gets stored

Inside each tenant brain:

• memories — text + 1024-d embedding + caution score + timestamp
• beliefs — text + confidence + SPRT state + supersession chain
• belief_evidence — every supporting/refuting event with note and timestamp
• predictions — prediction + domain + confidence + (when resolved) outcome + actual + surprise
• dreams — wander-produced thoughts + source memory ids + novelty
• agent_observations — multi-agent outside views
• tenant_activity — per-minute tool-call counter (for state)

Plain text. No PII required — Mneva does not ask for it. If you store PII inside the text of a memory, that's on you (and you should think hard about whether you want to).

What we never store

• The full API key (only the last 4 chars, as a display hint)
• Card data (Square tokenizes in the browser; we see only the nonce)
• Cleartext passwords
• IP addresses associated with content (nginx logs them for ops; they aren't joined to your brain data)
• Cross-tenant aggregations of memory content

At scale

SQLite-per-tenant is the current isolation model. At scale (tens of thousands of active tenants), this becomes schema-per-tenant in Postgres — the same contract, different physical store. The customer-facing API does not change.

Self-hosters can run the engine against the bundled SQLite or substitute Postgres; see Self-hosting.