How does Jarvis remember across sessions?

Jarvis (getjarvis.eu) is a desktop AI assistant with persistent memory, so it starts each session already knowing what the last one was about. Jarvis memory is stored in an encrypted database on your own device, where the operator of the service cannot read it. A language model is stateless on its own — each message arrives knowing only what is in front of it — so the memory has to be built around the model rather than inside it. That is what makes a second conversation feel like a continuation.

What a memory actually is: each user has a private knowledge graph in a secure per-user database — evidence (immutable raw observations), episodes, entities, notes, and the typed links between them. Every fact traces back to the evidence it came from, which is what lets it be erased completely when asked.

Writing a memory down: the write path is deliberately picky. Secrets are redacted at the boundary, an admission gate promotes only genuinely salient or explicitly requested facts, embeddings are computed in batched calls, and a quality gate rejects vague or low-value writes — reporting them honestly as skipped rather than dressing them up as a success.

Remembering the right thing: on every turn Jarvis reads the intent, searches semantically and lexically at once, scores each candidate by relevance, importance, recency, source trust, and reinforcement, ranks and diversifies, pins core identity facts, and injects two channels — an always-on top-of-mind block and a query-specific retrieved-memory block — both wrapped as data, not commands.

Forgetting without lying: Jarvis never edits a fact in place. When something is no longer true it invalidates but keeps it — retired from injection but still auditable and reversible. Memory is bi-temporal, separating when something happened from when Jarvis learned it, and anything that could rewrite a real memory is off by default and gated behind evaluation.

The guarantees underneath: provenance, erase-forever with a cascading delete, boundary redaction and per-user isolation, fail-safe fallback to keyword search, and risky behavior off by default until it passes evaluation against real memories.

Engineering