Nylogram: Accountable Memory for AI Agents

Nylogram is a Cartisien product. I designed it and I build it.

Agent memory is usually sold as magic: the assistant simply remembers. That framing works right up until someone has to answer a serious question about it. What did this agent actually know at the moment it made that decision. Where did that fact come from. Has it changed since, and who changed it. In a regulated industry or a vendor security review, "it remembered" is not an acceptable answer.

Nylogram is hosted memory for AI agents built around that accountability problem: model-agnostic recall with an exportable evidence trail, delivered as a REST API and an MCP server.

Memory is invisible infrastructure, and invisible things are hard to trust. The product had to make something inherently abstract inspectable without turning every developer into an archivist. Two audiences pull in opposite directions. Engineers want an API that stays out of the way. Compliance and security reviewers want to open the box and see receipts.

Provenance as a first class object. Every remembered fact carries where it came from and when. Recall surfaces that lineage instead of presenting a confident answer with no history behind it.

An evidence trail built for export. The audit view is designed around the moment it actually gets used, which is a compliance review or a security questionnaire, not idle browsing. It has to be legible to someone who was not in the room.

Change over time made visible. Memory that silently mutates is worse than no memory. The interface treats revision as something to be shown, so a user can see what a fact used to be and when it moved.

Developer surface and reviewer surface, kept separate. The API stays terse and unopinionated. The inspection layer sits alongside it rather than cluttering it, so neither audience is forced through the other's workflow.

Disclosure, provenance, confidence, and correction are the trust and control patterns my AI services are built around. Nylogram is those patterns taken seriously enough to be the entire product rather than a feature bolted onto one.