XTrace Memory API

XTrace is a memory API for businesses building AI agents. You send it conversation messages; it extracts structured memories, embeds them, and stores them in a per-org index you search with natural language plus filters. Three native types do the work: facts (a claim from a turn), artifacts (a doc, snippet, or summary worth keeping), and episodes (a session-level summary). You don't pre-classify anything. The server decides. One ingest call in, the right context back out. And it reasons instead of just retrieving: XTrace detects conflicts between sources, ranks what's authoritative, versions facts with supersede and retract chains, and knows what's outdated. Your agents get memory that tracks the current state of the world, not a pile of stale embeddings.

Procedural and shared memory

Here's what changes the math. When one of your agents figures out how to do something (underwrite a policy, triage a ticket, reconcile an invoice, run a multi-step workflow), that hard-won process is usually trapped in a single run and gone the moment it ends. XTrace's procedural memory lets agents share learned processes with each other, so one agent solves it once and every other agent inherits the solution. Now scale it. If your product runs agents for hundreds or thousands of customers, shared agent memory lets the whole fleet learn from every interaction in real time. A pattern discovered in one workflow makes every customer's agent better, while each customer's private data stays isolated in its own encrypted namespace. You share the intelligence, not the data. The result is a network effect on competence: the more your agents are used, the smarter every deployment becomes.

Encrypted by default

Memory is only useful if you can put real data in it. XTrace encrypts client-side with AES-256 and uses homomorphic encryption for vector search, so the server stores and searches over ciphertext and never decrypts it. Your secret key never leaves your environment, and we never train on your data because we can't read it. For vertical SaaS handling regulated customer information (health records, legal matters, financial transactions), this is the difference between a memory layer you can ship and one your security review kills.

Built to drop in

Wire it up in minutes. The TypeScript SDK is the primary client. Native Vercel AI SDK integration means you can wrap any model to auto-remember, or expose memory as tools your agent calls on its own. Async ingest by default for production loops, sync mode for scripts and demos, arbitrary metadata that becomes filterable on search.