Memento: memory that doesn't live in anyone else's cloud

I spent two decades building search engines so other people could store and find what they knew. Solr at Catho, then Semantix. The hard part was never indexing a document — it was deciding what counts as fact versus noise, and proving where every answer came from. That problem didn’t go away with LLMs. It got worse, because now it’s trivial for an agent to invent a memory with total confidence and zero evidence behind it.

That’s why we’re starting Memento today, February 26th. It’s a local-first memory engine for humans and for the agents working alongside them: it ingests your conversations, notes, documents, and Obsidian vault, and turns them into a memory you can actually query — with provenance, not vibes.

Local first, literally

Memento runs on your machine. Your data doesn’t go to a server of ours, or anyone’s, by default. That’s not a privacy marketing line — it’s the architecture. If you work with contracts, medical records, proprietary code, or anything that can’t leak, your agent’s memory can’t leak either.

I spent years selling — and building — enterprise search (Itaú, Magazine Luiza, Vivo, Petrobras) where the first question from any security architect was “where does the data live.” Memento answers that before anyone has to ask: on your disk.

Provenance over vibes

The second rule is uglier to engineer and easier to skip: every answer Memento returns carries where it came from. Which conversation, which file, which date. An agent that “remembers” something without pointing to a source is guessing, not retrieving. After two decades tuning relevance in Lucene, I know the difference between a good result and a convincing one — they are not the same thing. Memento is built to deliver the first.

Safe by default

Nothing is exposed to the network unless you decide to expose it. No open port by default, no telemetry shipping your content out, no “convenience feature” that quietly widens the attack surface. Safe by default means the easy path is the secure path — not an option you have to remember to turn on later.

What it already ingests — and why this matters for Arvor

Memento already imports Claude Code and Codex sessions, ChatGPT exports, WhatsApp conversation exports, Apple Notes, and entire Obsidian vaults. The idea is simple: your history of working with AI — and a good chunk of your life — already exists, scattered across half a dozen formats, each with its own structure, its own timestamps, its own mess. Memento pulls it into one place you control, one your agents can actually query instead of hallucinating context that never existed.

It’s the same lesson I learned indexing résumés and job postings at Catho, then tuning relevance at scale at Semantix: scattered raw data doesn’t turn into knowledge by itself. Someone has to design the ingestion pipeline, the normalization, the right index — and expect to get it wrong before getting it right. Memento assumes that from the design up.

That’s why Memento is the engine behind BRAIN MAKER, Arvor’s service for building personal and corporate brains — structured memories that real agents can use, with a source behind every answer. It isn’t another line on the roadmap. It’s the piece that decides whether an agent is useful or just another confident, wrong text generator.

We’re at the start of this project, launching today. What’s next is opening up more ingestion formats and making the provenance architecture auditable by anyone who wants to look under the hood. If you build agents and have felt the pain of a memory that “remembers” wrong, or want to talk about applying this inside your company, reach out through agentic consulting.