your ai remembers
what came before.
Ask about something you said weeks ago, revisit how a decision actually changed, or see what Backstory remembers. Everything is kept as versioned graph facts on HydraDB, not a buried transcript.
Graph native memory that reasons across long histories, changing facts, and multiple sessions.
real conversations
this is what it actually answers.
Every exchange around this is verbatim output from the running engine. Nothing here was written for the page.
how it works
every message becomes a graph fact, not just tokens in a transcript.
Backstory's engine runs the same five stage pipeline on every turn, backed by HydraDB as the graph of record. There is no vector only "closest chunk" guessing.
what it remembers
four ways long term memory actually gets tested.
Each of these is a real scenario you can run against the live engine, not a mockup.
questions
the things worth asking first.
What happens when a fact changes?
The old value is not overwritten. The new fact supersedes it, and both stay on the graph with their dates, so the current answer and the history behind it are both available. That is why the work question above can answer Company B and still show Company A.
What does it do when it does not know?
It says so. Retrieval can return related facts and the engine will still refuse if the specific thing asked about is missing, rather than answering from something adjacent. Refusing correctly is treated as a feature, not a failure.
Is my memory private to me?
Yes. Every account gets its own isolated graph, keyed to the signed in Google account and enforced on the server rather than in the browser. Nobody sees anyone else's facts, including the demo scenarios.
Why a graph database rather than a vector store?
Because the hard part is not finding similar text, it is knowing which version of a fact is current, what it replaced, and what contradicts it. Those are edges. On a vector store, superseded and current text look equally similar to a query.
Does it need an LLM to work?
No. An LLM improves how facts are extracted and how answers read, but the memory itself is graph operations. With no model configured the engine still stores, versions, retrieves and abstains correctly, and falls back automatically if a provider fails.