Why inra exists

Agents forget. Every session starts cold, and the usual fix — stuffing the entire history back into the context window — is slow, expensive, and gets worse the longer a user sticks around. Beyond a certain point you're paying to re-read a conversation the model has already had, and the one detail that actually matters is buried in the noise.

The other common answer — generic RAG over a document store — solves storage but not relevance. It returns things that look similar to the query, not the specific fact this agent learned about this user three sessions ago. Memory for an agent isn't a search index bolted on; it's a scoped, per-user, per-agent recall path that runs at query time.

inra is that layer. It stores what happened and returns only the memories relevant to the current step, keyed to the agent and user they belong to. Your agent reasons over the right context instead of all of it. Two calls — add and recall — and it works the same in a prototype or in production, behind whatever agent loop you already run.

We're a small team that ships. We'd rather show you working code than a manifesto, which is why the homepage is a quickstart and not a mission statement. The core is open (Apache 2.0) so you can read exactly what happens to your data, and the hosted service exists to save you from running a vector store and a recall pipeline yourself.

Skeleton page. Team, story, and contact details finalize with v1. Entity information above is verified and real.