Today's language models are extraordinary readers and extraordinarily forgetful colleagues. They can absorb a hundred pages in a second and lose all of it the moment the conversation ends. Within a single session they are fluent; across sessions they start over. And when they don't know something, the failure mode is rarely silence — it's a confident, well-formed answer that happens to be wrong.
We started Eidetikos because we think that gap is the central problem, not a rough edge to be smoothed over. Fluency arrived first. Memory has to arrive next, and it has to be the kind of memory you can trust.
The name is the thesis
Eidetikos comes from the Greek εἰδητικός — of vivid, accurate memory. We chose it because it states the standard we want to be measured against. Not memory as a longer context window, but recall that stays faithful to the evidence it came from, that can be updated when the world changes, and that is honest about its own limits.
A system that remembers accurately can tell you where an answer came from. It can say “I don't know” and mean it. It can be corrected once and stay corrected. Those properties sound modest, and they are the difference between a demo and something you would let near real work.
What we work on
Our research is organized into six areas that we think stand or fall together. Memory and recall is the core: grounded retrieval, persistent and editable memory that resists drift, and calibrated uncertainty. Interpretability is what makes that memory inspectable rather than merely asserted — mechanistic accounts of how representations form, and tools that surface why a model produced a given answer. Alignment and safety covers robustness under real-world pressure, oversight that scales with capability, and failure modes that are legible rather than silent. Grounded perception anchors language to documents, images, and structured evidence, so understanding is read rather than imagined.
Two more areas are less commonly named as research problems, and we think they should be. Frontier AI consumes enormous power and water to cool its hardware, straining the grids and communities where data centers live; we treat that as a first-class design constraint, which pushes us toward efficient inference and on-device execution. And cloud models routinely send personal data across the internet to be processed and, often, retained. We build so that sensitive information never has to leave your hardware to be useful — secrets in the system keychain, no training on private data without explicit, revocable consent.
The research pages lay out all six in detail.
Why a public benefit corporation
Nothing about the incentives of this industry naturally rewards moving carefully. So we wrote the obligation into the company structure instead of the marketing. As a public benefit corporation, Eidetikos is legally required to balance shareholder interests against a stated public benefit: the responsible development of AI for the long-term good of humanity.
In practice that means safety and societal impact get weighed in every significant decision — what we research, what we release, and how fast we move. When benefit and speed conflict, benefit wins, and we accept the cost of moving deliberately. It is a commitment we expect to be held to, which is rather the point of putting it in the charter.
What we're building
Five projects, building toward one system. Project Finch is the one we share openly: an evolving cognitive architecture built around long-term learning, exploring whether an AI can keep improving at your work over months and years rather than only within a single conversation. It is local-first, and it publishes its own progress — system health, devlog, and benchmarks — at projectfinch.com. You can watch it learn, including the parts that don't work yet.
The other four — Lodestar, Ursa Minor, Ursa Major, and Arcturus — we share by name only. The substance we walk through privately, under NDA, with investors and prospective partners. If that's you, request access.
What to expect here
We post when we have something real to share, which means this page will be quiet more often than not. Research notes, project updates, and the occasional announcement — published when the work warrants it rather than on a content calendar. We publish what advances understanding and hold back what could cause harm.
If you want to collaborate, join, or just ask a question, the door is open: get in touch, or see what we're hiring for.