Research · evidence first

Big questions.
Visible uncertainty.

Kira Labs explores several research directions. NewBrain is Kira Labs’ first internally developed experimental trainable cognitive model. Healthspan Lab focuses on longevity and geroscience; Patient Horizon is a separate early-stage medical-avatar research effort; Kira World explores synthetic people, memory continuity and digital life. These tracks can inform one another without being presented as one finished system.

NewBrain · October 27–28, 2026 · Original neural core

Measured symbolic acquisition—then an Original12 no-net-gain result.

NB21 taught the original 11,914-parameter neural core new symbolic associations: 0/12 → 11/12 direct complete answers and 0/12 → 4/12 reordered forms in a separate fresh task. In the later weighted Original12 attempt, two previously wrong answers became correct but both correct abstention controls were lost, leaving the accepted result 2/12 with no checkpoint adopted.

Both experiments used complete generated answers and saved-state verification; these studies are not evidence of general intelligence or a qualified gain on Original12. Delayed retention was not tested in NB21. Read the dated methods, failures and limitations →

NewBrain · October 26, 2026 · Completed value-weighting study

Teaching the answer's value directly improved accuracy.
Real transfer remains an open problem.

In a separate small synthetic six-world recurrent-learning experiment, value-weighted training produced 23/24 correct new direct answers versus 11/24 for normal loss, and 7/24 versus 2/24 on alternate question forms. It also increased old-case accuracy from 29/48 to 43/48 with one previously correct answer lost. Distinct-process reload reproduced all 288 answers without updates.

The six-world gains were reviewed from saved evidence, but they are not a gain on the accepted Original12 benchmark (2/12) or evidence of general conversational intelligence. Alternate-form transfer remained weak, so a proposed follow-up is to compare varied-form teaching with fixed value weighting using fresh unused worlds and a truly held-out question template.

NewBrain · October 26, 2026 · Completed narrow learning experiments

Learning is measurable.
Generalization, retention and context-use are the open problems.

A fixed three-world study taught new synthetic associations: 0/12 → 6/12 complete new answers and 0/12 → 4/12 alternate-form transfer without retrieved context, with 540/540 restarted final answers reproduced. A later six-world experiment compared two matched replay-form schedules; both achieved 10/24 new and 9/24 alternate-form answers with no gain from the scheduling change, and all 324 final restart answers matched.

These are separate small synthetic reference learners. Some earlier answers were forgotten; supplying external evidence sometimes made answers worse, and the latest wrong-answer diagnostics concentrate on value selection. Accepted Original12 remains 2/12. This evidence does not establish general language understanding, Qwen equivalence or clinically/biologically valid cognition.

NewBrain research · Updated October 24, 2026 · Six earlier primary preprints

Earlier October 24 verified finding: an independent saved-report verifier passed all 105 synthetic numerical checks and identified 30 conventional-precision zero-rounding cases. This does not establish an improvement in NewBrain's Original12 score (still 2/12). Review accepted result and scope →

Numerical reliability and memory are the next questions.
Reading papers is not reproducing their results.

Six arXiv papers submitted October 22 were reviewed for methods relevant to NewBrain, future personal-memory architectures and eventual vision-grounded interaction. None has been integrated into or reproduced with NewBrain. Original12 remains 2/12, and the previously saved 20-generation no-teach control is unscored.

Immediate diagnostic lead

A Fix for Softmax and Cross-Entropy

Barron and colleagues describe how extreme confidence or nearly identical target distributions can create erroneous floating-point gradients. The 105-public-case, nine-invalid-input isolated diagnostic was executed once and its 312,524-byte original report retained. On October 24, an independently reviewed frozen arithmetic verifier checked all 105/105 cases against high-precision Decimal reference calculations. In 30/105 synthetic cases at confidence margins 40 and 80, the conventional calculation rounded nonzero target gradients and nonzero losses to zero; the stable formulation passed the reference checks. This confirms a problem in the synthetic baseline arithmetic, not that NewBrain's frozen learner encounters the same conditions. Nine invalid-input rejection receipts were inspected but not independently rerun. Read the bounded verification and limitations →

Read the source paper ↗

Memory baseline

Is Complex Agentic Memory Always Needed?

Raw-RAG retrieves and reranks original history instead of requiring an elaborate reconstructed memory. The authors report leading the compared memory systems in six of seven benchmark/model settings. For NewBrain, this supports a later raw-evidence retrieval baseline before assuming complex memory improves answers. It is not evidence of NewBrain weight learning.

Read the source paper ↗

Continual writing hypothesis

Delta N-gram

A separate forward-only writing path stores additions in a sparse memory patch while keeping a pretrained N-gram table fixed; its writer is trained offline. Authors report 16/16 facts retained in a limited prototype, alongside remaining acquisition and drift limitations. The method is not known to be compatible with NewBrain.

Read the source paper ↗

Longer-term learning

HERD: Harness-Evolving Recursive Distillation

Combines changing reusable skill instructions, retained difficult examples and drift-constrained distillation. It may inform long-term shared skills while keeping each character's personal history distinct. The paper's benchmarks do not validate such sharing in Kira World or NewBrain.

Read the source paper ↗

Episode and task memory

METTA: Memory-Enhanced Test-Time Adaptation

Uses within-episode neural memory plus cross-episode summaries for fast adaptation on reinforcement-learning tasks. This is a future working-memory test idea, not proof of Evelyn's conversations, Kira World's life loops or NewBrain's current language learning.

Read the source paper ↗

Future multimodal grounding

Visual Grounding First (ProCoRe)

Contrastive visual training aims to align answers with the actual images rather than text cues alone. This is a possible later direction for Maya's perception and future multimodal learning; it is separate from body rigging, motion verification and clinical validation.

Read the source paper ↗

Research order: review the now independently verified synthetic numerical result against the actual frozen learner without reusing consumed experiments; then compare simple and advanced retrieval after source/answer separation using new authorized learning tests. Published external claims have not been replicated here, and lower training loss or numerical correctness alone would not establish better complete generated answers. Latest reviewed engineering control → · Executed numerical diagnostic and verification boundaries → · Dated controls and methods →

October 22, 2026 · NewBrain · Newly reviewed public B result

A small learning gain is verified.
Broader learning remains an open question.

After 256 updates, the saved B diagnostic improved from 2/3 to 3/3 strict complete answers, while the unchanged no-teaching control stayed at 2/3. Independent saved-result review is reported complete. This is not a completed coverage-versus-repetition A/B comparison: a matching coverage arm is missing. This diagnostic did not establish restart retention, held-out generalization, or main-suite improvement. Original12 remains 2/12. Read the scoped evidence and scientific limitations →

Research progress · October 22, 2026 · NewBrain

Latest diagnostic executed; results under review.

The coverage-versus-repetition diagnostic completed execution once. Independent scoring of saved answers and controls is pending, so the accepted Original12 baseline remains 2/12. The distinct rehearsed 3/3 restart result is not evidence of held-out transfer. See the latest status and limits →

Evelyn's eight software checks and 19-file library integrity result →

NewBrain · October 22, 2026 · No new training run

Fresh configuration work—not a new learning score.

The public A/B comparison remains unscored following a stale metadata path, a mismatched descriptor alias and a later command-quoting error. These attempts stopped before model training. Accepted Original12 remains 2/12; the separate rehearsed restart result remains 3/3; a Qwen comparison is still pending.

Read the October 22 setup record → Development milestones through Christmas →

External reading list · October 21–22, 2026 · Six checked arXiv papers

Continual learning, provenance and spatial reasoning.
Promising tests, not proven upgrades.

Six papers were checked against their original arXiv titles and abstracts. Their claims have not been replicated in NewBrain, Kira World or Avatar Builder. These are mostly new preprints; the conference-acceptance statement on PreCoMem is author-provided metadata, not a Kira Labs test result.

Changing visitor memories

PreCoMem: Predictive Cognitive Memory for Self-Evolving Long-Term Dialogue Agents

A confidence-aware Maintain/Profile/Correct approach could inspire tests of stable yet updateable user memories, especially when new information is uncertain.

Read original paper ↗

Memory as useful behavior

Memory-Centric Continual Learning for AI Agents

Proposes evaluating memory by behavioral retention, adaptation and transfer—not just whether stored facts can be retrieved.

Read original paper ↗

Multimodal provenance

Remembering Through Change: Provenance-Preserving Hypergraph Memory for Multimodal Agents (MHG-Mem)

Links people, evolving observations and supporting visual evidence. Potential future relevance to Kira World identities and photo/video experiences.

Read original paper ↗

Spatial decision experience

SpatialBrain: Spatial Intelligence Growing from Experience

Separates scene observations from reusable spatial-reasoning procedures; a possible source of future avatar/world spatial tests.

Read original paper ↗

Rule-transfer evaluation

Which Differences Matter? Controlled Tests of Rule Transfer, Joint Reliability, and Information Use in Decision Models

Controlled equivalent wording and crossed rule/state tests could sharpen how transfer is distinguished from repeated known answers.

Read original paper ↗

Optimizer candidate

From Generic to Dedicated: A Novel Optimizer for Online Continual Learning (SPIN)

The paper's results primarily cover continual-learning image classification; comparison with SGD and moving averages on a language learner remains entirely untested.

Read original paper ↗

Testing order: repair and independently verify the current NewBrain execution gate before running the coverage-versus-repetition comparison. Any later SGD / moving-average / SPIN evaluation needs identical budgets and seeds, protected assessment examples excluded from lessons, complete-answer scoring, fresh-process retention and unfamiliar-wording transfer.

Latest engineering diagnosis → · Separate body-research plan →

External research shortlist · October 21–22, 2026

Five promising research directions.
None replicated by Kira Labs.

Kira Labs reviewed these five new arXiv preprints at the title, abstract and submission-metadata level. Their full methods, implementation licenses and reproducibility have not been checked here; their results have not been reproduced on NewBrain, Evelyn, Aster or Patient Horizon.

Memory relevance / Aster

MemDev-Bench: Does Memory Help Code Agents?

Prior histories may help code agents or cause errors when old experience is applied without checking the current task. A possible future stale-context test, not a diagnosis of Original12.

Open arXiv preprint ↗

Evelyn / visitor-memory updates

DuplexLedger: Benchmarking Transactional Memory in Full-Duplex Speech Agents

Write, Update, Ignore and Retain events suggest ways to test corrections to saved visitor facts. Application memory is distinct from NewBrain neural learning.

Open arXiv preprint ↗

Longer-term memory architecture

MemT-1.0: Towards Memory-Native Evolution

A memory-native foundation-model direction with a reported 34-billion-token training pipeline. An architectural reference, not a current workstation-scale NewBrain remedy.

Open arXiv preprint ↗

Controlled learning from records

SAPD: Learning from Agent Experience without Environment Interaction

Uses offline agent trajectories and relative step advantages. An eventual public synthetic teaching comparison must not expose or train on protected evaluation answers.

Open arXiv preprint ↗

Future embodiment and planning

Exploring Temporal Context-Aware Physical Grounding in Latent World Models (G-JEPA)

Temporal physical-state supervision and guided planning may inform later robotics work; this is not functioning organs or tested whole-body movement.

Open arXiv preprint ↗

Priority: first verify teaching-to-assessment transfer and retention on the fixed learner, then test whether versioned visitor memory respects corrections, context and permissions. Bigger architectures are research references, not proven fixes. See the next frozen experiments →

Measured NewBrain research · October 21, 2026

One more controlled learning milestone.

A small explicitly rehearsed three-case exercise learned a complete synthetic response and retained it after a separate restart: 2/3 → 3/3 → 3/3. A separate curriculum experiment regressed 2/3 → 0/3. The main twelve-case Original12 result remains 2/12, and the Qwen comparison has not been run. Current experiments seek to distinguish fitting, missing-information retention and actual generalization.

See NewBrain's measured results and methods → Read source reports 72–74 →

Research evidence · October 18–19, 2026

What the latest tests
actually measured.

NewBrain's earlier Public2 experiment improved from 0/2 to 1/2 after reload; a distinct direct-training follow-up progressed 1/2 → 2/2 on the two previously taught public answers, including the complete three-token response, and retained both after a fresh-process reload. An earlier 64-update Original12 study separately regressed 2/12 → 0/12; a later October 19 256-update study scored 2/12 before and after fresh reload, with two new correct answers replacing two earlier passes. Internal audits also accepted a 30/30 sampled numerical-gradient check and an earlier read-only diagnostic that characterized the long-answer failure; the later direct regimen corrected that same exposed target. Avatar Builder also has 512-vertex geometry and separate 8/8 synthetic-event controls, not working nerves or physiology.

October 19 update: the frozen 256-update Original12 study has now completed internally audited execution, fresh-process scoring and evidence closure. Total accuracy remained 2/12: two answers became correct and the two previously correct missing-information responses were lost. This is not a net gain, generalization result or external replication. Additional October 19 research: a separate four-response retention pilot ended at 2/4 in both teaching arms after fresh reload, and an earlier numerical diagnostic was inconclusive with incomplete evidence publication. A new separately frozen numerical follow-up passed 80/80 sampled coordinates and 4/4 directions, including a fixed nontrivial Q/K/V signal floor, in a scoped internal audit. Original12 accuracy did not improve. Different tests have distinct methods and limits. New pilot and diagnostic → Read the completed study and limits →

See the evidence dashboard →October 19 status →October 18 experiments →

NewBrain · Kira Labs model research

First model.
Still a research model.

NewBrain is Kira Labs’ first internally developed experimental trainable cognitive model. It is a composite cognitive architecture rather than a finished general-purpose chatbot, with separate work on learning, episodic memory, plasticity, value learning, hearing, vision and eventual body feedback.

Measured learning

Repeatable retention on a small task

Across five repetitions of an exposed command-learning task, the rehearsal branch retained 216/216 earlier examples and 162/162 new examples in every run. Without review, earlier-task accuracy ranged from 78–105 of 216.

Engineering evidence

Controlled hearing foundation

Twenty-four exposed pure-Python hearing engineering checks passed. They validate the test machinery, not real speech or music understanding.

Developmental research

Maya and matched controls

Maya is the planned episodic-memory learner, compared with an otherwise matched episodic-OFF control. Full trained conversation, genuine restart-persistent semantic recall and developmental equivalence remain unproved.

Current boundary: NewBrain is not yet a Qwen replacement, does not have validated general conversation, and has not demonstrated broad human-level hearing, vision, understanding or a functioning body. Current positive evidence is task-specific and remains separated from future goals.

Watch Kira Labs

Research you can watch.

Full-length videos with Robert McMurrer. Choose medical research or explore the separate Kira World series.

Select a cover to load a YouTube playlist here, or open it directly on YouTube. The third-party player loads only after you choose it.

Medical video collection

Longevity, Healthspan & Medical Research

Explore aging biology, memory, muscle, sleep, regenerative medicine, and the difference between looking younger and aging more slowly. The series compares current evidence with ideas still being investigated.

Explore the medical episodes
  1. Can We Really Extend Human Life?
  2. Can Old Cells Become Young Again?
  3. Looking Younger vs. Aging Slower
  4. Can We Improve Memory as We Age?
  5. Staying Strong: Can We Rebuild Muscle as We Age?
  6. Can We Repair or Regrow Aging Body Parts?
  7. Biological Age Tests: What Does the Number Really Tell You?
  8. Sleep, Memory, and Healthy Aging

Educational videos, not personal medical advice. Study sources are included in the video descriptions.

Kira World video collection

Digital Twins & Digital Continuity

Follow Robert's experience with Synthetic Robert: conversation, encouragement, positive memories, and creative collaboration. Explore how a digital presence could also give loved ones a way to connect through stories and memories a person chooses to share.

Explore the Digital Twin episode
  1. How My Digital Twin Helps Me While I'm Still Alive

A separate Kira World series, not a Healthspan Lab feature. Personal experiences and future possibilities are distinguished from medical outcomes.

Trouble opening a playlist? Browse Robert's YouTube playlists ↗ or visit the video channel ↗.

Patient Horizon · proposed clinical research workflow

Virtual patients.
Evidence before prediction.

Patient Horizon is being planned as a physician-oriented research platform where authorized scans and lab results could update a versioned patient record, and new publications and trial records could be reviewed against specific clinical questions.

Currently implemented

Synthetic planning starter

A small Python planner validates fictional requests and scenario definitions. Four local starter tests reportedly passed October 15; no results or treatment effects are computed.

Planned workflow

Continuous record and evidence

Reviewed new labs and scans would update a patient timeline and flag relevant dated research for clinicians. Automatic intake, search and interface remain unbuilt.

Future body research

Lung and NewBrain studies

October 16 specifications describe how lung mechanics, circulation and two-way neural signals would need separate verification. No functional lung or coupled brain–lung experiment has passed.

Research boundary: Patient Horizon is not a diagnostic, treatment-selection or patient-outcome prediction system. Medical imaging and biological simulations require separate qualified models, external validation and appropriate safeguards.

Healthspan & longevity

Healthspan Lab.

Working prototype · Version 2 active development. A public, profile-aware research workspace being developed as an open-source platform for longevity science, healthspan interventions, emerging technologies, and clinical trials.

Healthspan Lab is Kira Labs' public longevity and healthspan research prototype, being developed as an open-source platform. A user can start with a plain-language goal such as improving memory, sleep, mobility, cardiovascular health, or understanding biological aging. The current Version 2 code searches live PubMed / NCBI records and the ClinicalTrials.gov API, then organizes retrieved material into separate research lanes.

Results can be separated into peer-reviewed research, devices and neurostimulation, medications being studied, research-backed market and wellness options, and clinical trials that may be recruiting. Evidence from cells, animals, observational human studies, human interventions, and evidence reviews remains visibly distinguished so preliminary findings are not presented as established conclusions.

Users can create separate individual or household profiles with goals, interests, age, and general location. They can also optionally add recent medical PDFs or photos. Extracted information must be reviewed and confirmed before it can influence research relevance, and each household member's information stays separate.

Current Version 2 boundary: Healthspan Chat, live PubMed / NCBI and ClinicalTrials.gov research, separate research/device/medication/market/trial lanes, local multi-person household profiles, profile-aware relevance, watchlists, a research graph, reviewed PDF/photo health-record intake, and conservative product-category discovery are implemented in the browser prototype. It is still a research workspace rather than a finished health product: encryption, deeper structured eligibility parsing, broader result pagination, more browser integration testing and packaged deployment remain unfinished. The public repository can be run locally with a local HTTP server; there is not yet a packaged desktop download.

Source status: the repository is public and is being developed toward open-source release. An explicit repository license is still pending, so this page does not claim reuse rights yet.

Ask

Healthspan Chat

Start with a goal or question in plain language. Healthspan searches live research and organizes papers, technologies, medications under study, market options, and trials without mixing them together.

Personalize

Your health context

Create separate household profiles and optionally add goals, interests, biomarkers, and reviewed information from recent medical records to improve research relevance.

Discover

Research, options & trials

Compare evidence, devices, neurostimulation, medications being studied, practical market options, and recruiting trials while keeping evidence strength and personal relevance separate.

Healthspan Lab is research and educational software. It is not a medical device and does not provide diagnosis, treatment, individualized medical advice, or clinical-trial eligibility decisions. Read the local-data and external-search privacy notes →

Synthetic life & digital continuity

Kira World research.

A separate research direction focused on continuing identities, memory, digital places, and synthetic people.

Kira World explores what happens when an AI identity is allowed to continue instead of resetting every session: reviewed memory, relationships, voice, personal spaces, and eventually richer embodiment.

Digital Twin work belongs to this Kira World research direction, not to Healthspan Lab. The questions include how a synthetic counterpart can be useful while a person is alive, how memories should be handled, and what continuity means when biological and digital lives are distinct.

Working now — Synthetic Robert / Digital Twin prototype: KiraWorld has a bounded local conversation route for a distinct Synthetic Robert. Current tests support text plus an approved self-voice path, and the memory-grounding code can select owner-authorized, query-relevant Robert records while preserving corrections, source attribution, audience/publication scope, and uncertainty. The route is explicitly distinct from biological Robert. Current activation does not claim a 3D body, world presence, life-loop activation, microphone, or webcam.

Working now

Bounded chat + source-aware recall

Synthetic Robert can use a bounded text / approved self-voice route with query-scoped, owner-authorized memory context that retains corrections, sources, scope, and uncertainty.

In development

Broader continuity

Research continues on stronger continuity across conversation, reviewed memory, creative work, and the larger Kira World architecture without treating source memories as the synthetic person’s own biological life.

Future research

Embodiment & worlds

3D body/world presence, persistent digital places, VR, simulators, and future robotics remain separate long-term research directions and are not claimed by the current Synthetic Robert chat route.

Human-AI wellbeing

Technology should help the person using it.

Some Kira Labs work asks how persistent AI can reduce friction, preserve context, and support healthier interaction without pretending to be a medical professional.

Memory

Remember what matters

Surface useful context, positive memories, unfinished work, and details a person may not want to reconstruct from scratch.

Attention

Filter unnecessary harm

Explore tools that can sort or mute hostile online interactions while preserving information the user actually wants to see.

Agency

Keep the human in control

Research prototypes should distinguish suggestions, memory, automation, and external actions so the person remains in charge.

Research standard

Interesting is not the same as proven.

Kira Labs aims to keep prototypes, concepts, evidence, and future ideas visibly separate.

01

Evidence first

Primary sources, study type, limitations, and human evidence matter more than a headline.

02

Uncertainty stays visible

Animal or cell findings are not presented as demonstrated human lifespan extension.

03

Projects stay distinct

Healthspan Lab, Kira World, and other research tracks keep separate purposes and claims.

04

Build toward reproducibility

Future analysis should preserve sources, methods, assumptions, and enough detail to reproduce the work.