Blog
Our values, project notes, product work, and reproducible head-to-head tests of mainstream AI models. Every test's inputs and outputs can be reproduced.
What a Personal AI Needs Instead of Just a Bigger Brain
The Person-Centered Series 07: devices change, so what stays continuous? Not a bigger model — a personal relationship world model, one portable model of who a person is, their people, and their history. Why that is a different layer from embodied intelligence, and what breaks without it (MEMENTO measured a 30.5% drop in task success on frontier models).
2026-10-30 Product NotesWe're Not Competing with Your Memory System
The Person-Centered Series 06: the positioning post. Engine, storage, subject: three layers, three problems, and we're building the one that answers whose memory it is. Model-agnostic by design, additive to the stack you already run, no retraining. Plus an honest read of what ChatGPT Memory-style automatic profiling gets right.
2026-10-26 Self-EvolutionWhen the Judge Can't Judge: U Is a Verdict, Not a Failure
Self-Evolution Notes, entry 03: the LLM judge in the middle of an automated loop will eventually fail. Crash, fake fallback, or a first-class undetermined state. How we built the degradation path, the audit that caught mislabeled verdicts, and the self-tests that keep the judge honest.
2026-10-22 Embodied AIRobots Need to Know Who You Are
A house has four people and one account. Face recognition tells a robot who is in the room; it cannot tell it whose preferences govern which room. Embodied memory research solved how (Voyager's skill library), what worked (ExpeL), and where (3DLLM-Mem), but left out whom. And why provenance becomes a safety property once memory has a body.
2026-10-19 ValuesThe Next Memory Wave Won't Be Bigger: It Will Have a Subject
The field is racing to make agents remember more. We think the next wave is about the subject, not the capacity. Four pressures pushing that way: the benchmarks are starting to measure the person (PersonaMem: frontier models around 50%), pure capacity has a visible failure mode, sovereignty rights need a person root, and the second person is already in the house.
2026-10-19 ValuesEveryone Is Unique. That's Why We Won't Build a Generic Memory
Every AI company is selling memory that stores the same things about everyone. Our position: your world is made of particular people, commitments, and history, and it belongs to you. The value behind everything we ship, stated as an acceptance test.
2026-10-19 ProjectSecond Look: A Second Opinion on the Advice You're Getting
Advice is usually well-meant. The risk isn't. Second Look breaks a piece of advice into three qualitative signals and closes with one question you can check: for this to be good for you, what has to be true? No motive-reading, no trust score, no deciding for you.
2026-10-19 Project NotesWhat Helm AI Is Building: Two Tools, One Principle
New here? Helm AI builds practical tools that treat AI output as a draft, never a conclusion. Two tools (resume rewrite, writing assistant), one project in validation (Second Look), and the single principle they share.
2026-10-19 ProductWhy Our Resume Rewrite Can Promise Not to Fabricate
The deadliest error in AI resume rewriting isn't grammar, it's quiet inflation: verb upgrades, invented numbers, widened scope. Here's the second AI pass that checks every bullet against your original, and a three-step test you can run on one free bullet.
2026-10-19 Self-EvolutionYour AI Remembers What Happened. It Can't Remember Who You Are.
Self-Evolution Notes, entry 02: your AI recites everything you told it, but has no idea who you are. A full read-only audit of our own memory layer found that 46% of personal content can't be resolved to a specific person. The problem isn't the model. It's the shape of the storage, and why "don't guess" is the first rule of person-centered memory.
2026-10-18 Self-EvolutionVersion Your Judge First: Notes from Building a Self-Evolution Engine
Opening entry of Self-Evolution Notes: we're building an engine where humans set the goal and the system finds the path. Lesson one comes from a real incident in the evaluation layer. An LLM judge that drifts is a loose steering wheel. Three anti-drift rules (version the judge, batch the weights, keep feedback away from the gates), plus how one incident became discipline.
2026-10-18 ValuesWe Studied Every Major Memory System. None Can Be the Base Layer for Your World
Mem0, Zep, Letta, Cognee, ChatGPT Memory, MemoryBank, A-MEM, HippoRAG, Generative Agents: four families, nine systems, six dimensions, and five structural gaps that keep any of them from holding a person. Eleven arXiv citations, all checkable.
2026-10-18
Join the discussion
Feedback on the blog itself, the series, or a topic you want covered next? Leave a note below. If you are working on something person-centered, or you think we are wrong, write to support@helmaiagent.com. Individuals and organizations interested in this direction are especially welcome.