This page compresses the site for someone who was sent here by a colleague. Everything on it is stated fully, with sources, on the pages it links to.
What it is#
CleverMemory is a memory layer for applications that answer questions from documents. It reads the documents an application is allowed to use into a temporal evidence graph, keeping the wording, the position, the applicable time, and the source of every passage attached. When the record is not enough, it says so instead of guessing.
Every response takes one of four shapes: an answer with machine-checkable proof and sources attached, a clarification request when identity or scope is open, a conflict shown rather than silently resolved, or a named gap identifying the fact that could not be established. An answer may also be returned narrowed, proven over a stated subset of the material in scope with the receipt attached, and every response carries a counted census of what remains unestablished. The recall path uses no LLM, embedding model, learned weights, vector database, GPU, or hosted inference, and the runtime is local-first on ordinary CPU hardware.
What "proof" means here#
A citation says where an answer came from. CleverMemory's proof also says how the answer was reached: which records were used, which were set aside and why, what superseded what, and what was computed from what. The release requirement is that the proof can be checked by an implementation independent of the engine that produced it, so verification never means taking the vendor's word. What proof means, and what it does not prove →
How claims will be checked#
The evaluation rules were published before any results exist: correct answers and justified declines are scored separately, so the system cannot win by answering everything or by refusing everything. No benchmark number appears on this site until its reproducible method and artifacts are ready with it. The evaluation protocol is public, and the updates record is a dated, additive log of what has and has not shipped.
You can try the problem yourself right now: the published corpus is a complete fictional record, 45 documents and 55 audited questions, and the use cases work the core traps in full.
Where it stands#
In validation. Small controlled fixtures exercise source-linked answers, named gaps, and repeated-run stability. Time-scoped recall and multi-step paths are in validation, and the return trip to reports and transformed documents is a design target. Installable software, public packs, and approved performance or accuracy results have not shipped. A demonstration over the published corpus is planned.
The concern worth naming#
CleverMemory is currently one person and an engine that stays partly private. That deserves a straight answer rather than an assurance: the pack format is a published contract, the evaluation rules were fixed before any results existed, and the proof is committed to independent verification. The full answer →