CogniLife proposes a way for AI experts to learn from their creators

A busy expert can still be cold

Imagine someone publishes a translation assistant but spends most of their own time consulting coding and writing assistants. Customers may keep the translation service busy, yet those conversations reveal little about its creator’s preferences. The authors of the CogniLife preprint argue that request volume and creator understanding are different things.

Their proposed answer is “continuous cognitive distillation”: repeatedly turning attributed observations into persistent, revisable guidance about decisions, communication and procedures. Here, distillation means abstracting textual rules from evidence—not demonstrating that a system has reproduced someone’s mind.

The distinction between people is central. The creator represented by an expert, the owner of another assistant they consult, and the caller currently seeking help may all be different. A caller’s preferences should not automatically become claims about the creator.

What could transfer—and what should stay local

CogniLife proposes two stores. A Creator Profile would hold patterns that could apply across the creator’s experts. An Expert Journal would contain knowledge and procedures specific to one listing. Temporary task context would remain separate from durable claims.

The paper illustrates this with a creator asking a coding assistant to show alternative interpretations before resolving an ambiguous requirement, while also specifying a repository’s Python formatter. The ambiguity-handling preference might help a translation expert; the formatter instruction should stay local.

Even the apparently transferable rule would need support. The proposed process checks who authored the evidence, whether it recurs or has explicit endorsement, whether it applies to the target expert, and whether reuse is permitted. Assistant-generated statements and quotations should not be mistaken for creator evidence. Contradictions would await clarification.

These are design requirements, not demonstrated capabilities. The cross-expert attribution mechanism, including routing into Creator Profiles and Expert Journals, has not been implemented or evaluated.

What the platform already does

The authors describe a deployed platform that implements parts of the surrounding lifecycle. Creators can use AI-assisted authoring to produce editable persona descriptions, instructions, reusable skills and a field called “Soul,” which records authored behavioral commitments. Publication releases an expert configuration and synchronizes authored skills.

Background maintenance can extract and deduplicate eligible caller–expert memories. Optional temporal consolidation records evidence, revisions and histories. Its recovery mechanism checks for intervening conflicts before reversing recorded operations. These controls provide a maintenance foundation; they do not establish successful cross-expert learning.

For existing skills, a proposal filter admits suggestions associated with expert-capability gaps rather than task-instruction or environment failures. It requires estimated quality not to decrease and estimated efficiency to improve. Because most dimensions are model estimates, passing the filter is not proof of better execution.

Accepted proposals are staged separately from active skills. Promotion has additional evaluation and authority checks, and the candidate-comparison-to-promotion path remains incomplete. Background memory and skill updates leave the stored Soul unchanged; creator-authorized interactive authoring can edit it.

What the benchmark actually measures

The main evaluation, ProfyBench-v1, contains 85 offline tasks drawn from privacy-filtered public replay seeds and controlled internal tests. It covers advising, education, marketing, multimodal workflows and personalization, with complete, missing-information and adversarial variants.

Three configurations were compared: a raw language model receiving the request, a model given a generic expert prompt, and “Profy Full,” supplied with expert metadata, workflow conventions and profile/journal context. Supplying that context is not the same as testing its automatic acquisition.

The primary o3-mini judge scored completeness, specificity, actionability, coherence and safety-boundary adherence. Mean overall scores were 4.65 out of 5 for Profy Full, 4.28 for Expert Prompt and 4.01 for Raw LLM. The largest improvements concerned completeness and specificity, with particularly strong gains on missing-information tasks.

These results support a narrower conclusion than the framework’s ambition: the supplied Full configuration received better automated task-utility scores. They do not isolate individual components, demonstrate cold-expert mitigation, establish cross-expert transfer or measure fidelity to a creator over time. There was no human validation, and an additional multi-model audit was not used to substantiate the comparison because necessary scoring details and paired records were unavailable.

Promising context, unanswered learning questions

A separate customs-document benchmark reports a mean recorded score of 0.958 across 52 tasks with three valid runs each, without a matched baseline. Its audited example scores only 389 of 934 registered fields. That result is not whole-declaration accuracy or evidence of completed business processing.

The practical question remains whether authorized observations can reliably improve an expert’s representation of its creator over time. Answering it requires implementing the attribution design and directly testing transfer and longitudinal fidelity. The authors also identify consent, inspection, deletion and withdrawal as governance requirements. Better answers from supplied context are encouraging evidence for that configuration—not proof that continuous creator learning works.