
Independent research for real-world impact
Research today.More capable tomorrows.
Kira Labs builds experimental AI systems that study learning, memory and continuity. Our projects connect cognitive research, synthetic people, medical-avatar studies and future robotics—with measured results and clear limitations.
Kira Labs update · November 8, 2026
Hardware has slowed NewBrain.
It hasn't stopped us.
We're working toward more RAM and computing capacity before larger NewBrain training experiments can continue. But Kira Labs isn't standing still. We're developing Kira World environments, character-creation tools and Avatar Builder research while the hardware catches up. Recent work includes an early, keyboard-controlled Louvre-inspired walkthrough draft—not a finished museum recreation or a populated virtual world. Watch the five-minute Louvre V66 walkthrough → · View Home World and Louvre screenshots → · November 9 research update →
If you'd like to help, please speak with Robert, visit our Support page or explore the Kira Labs Shop. Equipment, technical advice, collaboration and sharing the work can all make a difference.
New film · October 26, 2026 · 3 min 12 sec
Watch Kira Labs.
Research today, more capable tomorrows.
Evelyn narrates this completed overview of Kira Labs, verified component work, NewBrain research and ways to support future experiments. Actual demonstrations, licensed development footage and concept illustrations are clearly distinguished.
Our accepted Original12 score remains 2/12. The original Evelyn head and complete body are not yet finished.
October 25, 2026 · Actual recordings and measured research
Watch Evelyn speak.
See NewBrain's learning tradeoff.
Evelyn's real 58-second developmental recording contains four locally generated replies in her approved synthetic voice, with scripted head and eye movement, blinking and speech-timed facial animation. This is a reviewed offline replay with a licensed developmental model—not her final Avatar Builder body or autonomous motor control.
A separate five-world NewBrain experiment showed stronger retention but weaker composition than blocked training (14/40 old cases and 12/20 composition). Accepted Original12 remains 2/12.
Our research programs
Different paths. A connected vision.

Learning and memory experiments, including October 26 new-fact and question-form gains, with unresolved retention and value selection.
Explore NewBrain →
Medical record and digital-anatomy research with planned clinician-reviewed applications.
Explore Patient Horizon →
Continuing people, personal apartments, research spaces and distinct memories.
Explore Kira World →
An emerging individual with approved original voice and recorded replies; final body, open-ended dialogue and release testing remain ahead.
Meet Evelyn →
ROS 2 and embodiment research, with simulated safety gates before any real-world robot.
Explore Robotics →
IdeaForge, Video Studio, HumanoidResearcher and other practical research projects.
Explore Tools & Apps →About Kira Labs
Evidence first.
A more capable and humane future.
Kira Labs is an independent AI research initiative based in Newark, New Jersey. We explore cognitive architectures, persistent synthetic identities, accessible technology and future medical and robotic applications—while documenting what works and what still needs testing.
Research evidence · reviewed November 7, 2026
Measured results.
Clear limitations.
The latest completed NewBrain language-learning study failed all four outcome categories despite exact restarts. Avatar Builder now exports changing animation clips, but its original human anatomy remains unapproved; Aster has separate limited synthetic retention research, not a conversation-ready brain. NewBrain’s protected Original12 remains 2/12.
Browse downloadable reports, test results and real images →November 7 · NewBrain and Avatar Builder: Fresh locked Native research trials improved different tasks without a complete pass: V11 supported 8/8 and combined 4/4 but abstained only 1/5; V14 abstained 5/5 but lost other accuracy. V95 Avatar proportions respond to silhouette references, but original anatomy is still unapproved and two Avatar unit tests errored. Read the dated research →
November 6 latest · Avatar Builder V77: The original-body export now contains five changing animation clips (129 passing tests, one skipped). Anatomy is unchanged from a rejected draft, and independent validation still found failures; no accepted Evelyn body or Native NewBrain learning gain. Read the verified update →
November 6 later · NewBrain and Avatar Builder: V68 repaired small audio and memory/tissue software controls, with 84 scoped NewBrain checks plus 5 independent controls and 36 bounded Avatar Builder checks. These are component results, not successful native learning or an accepted full body. Read the newest research summary →
November 5 later · NewBrain and Avatar Builder: A separate small decoder retained taught replies after reload and benefited from rehearsal, but failed unseen unknowns and new wording; it is not a repaired native NewBrain. An original-head fitting branch advanced through photo/form fit before failing eye clearance. No accepted complete face, body, or Original12 improvement. Read the new research summary →
November 4 later · NewBrain and Avatar Builder: the latest saved NewBrain V9 learning study is active but unscored, and a first-answer credit diagnostic is proposed. Original head V8 and full-body V5 produced real rejected/failed results; saved geometry points to early jaw/neck control-mesh distortion, while corrective head V9 and body V6 sources remain unrun. No new accepted model or human body. Read the current engineering findings →
November 4 later · NewBrain V8v2: a reviewed 294-member checkpoint/source snapshot and resource calculations are preserved, but there is no measured full-process RAM fit or new learning score. The November 2 negative study and Original12 2/12 stand. Read the qualified scope →
November 4 later · Avatar Builder V6/V7: V6 failed a geometric projection guard; V7 passed source review only and has not been rendered. The blue eyes are still preserved, while the head/body remain rejected. Read the failed run and next steps →
November 4 · NewBrain checkpoint memory: compact saved-state and buffer-lifetime repairs passed limited synthetic tests, but full native execution, real-process RAM savings and improved learning remain unverified. Read the research limits →
November 4 · Avatar Builder head V5: one actual original head built and rendered 13 static views (12,680 skin vertices, 25,072 triangles). Approved blue eyes remain intact, but the head is still rejected due to facial anatomy. Read the latest visual review →
November 2 · NewBrain scored learning: a fixed native readout study completed 216 training updates and 72 generations. Taught facts, paraphrases, unfamiliar two-fact use and unsupported-question abstention each scored 0/6 before, immediately after and after restart. The 24/24 exact restarted replies were still wrong. Protected Original12 remains 2/12 and was not retested. Review the evidence →
November 2 · Avatar Builder workspace: launcher identity validation passed 12 bounded tests and Windows HTTP checks, but no newly accepted human head, body, mesh or movement followed. The previously approved round eyeballs remain preserved. Read the actual scope →
Aster · October 25 study, integrated November 1: a separate synthetic interleaved-replay experiment retained 59/72 older answers versus 50/72 with blocked practice, but all compound-format and unseen tests failed (0/60 and 0/24). This is not general conversation. Read the bounded Aster review →
November 1 · NewBrain optimizer candidate: an isolated, opt-in Adam update path now has narrowly qualified optimizer-state adoption and fresh-process save/reload checks. It is software persistence progress, not better language learning; incomplete training and held-out behavior remain unproven, and Original12 is still 2/12. Review the actual scope →
November 1 · Avatar Builder source update: a new original profile-driven exterior field supports 68 explicit shape controls and source pathways for 2D reference guidance, guarded native meshing and request dispatch. Small synthetic fixtures do not prove a finished body: real native execution, likeness, animation and internal anatomy remain unaccepted. Read the new development boundary →
November 1 · NewBrain historical memory: an isolated, independently checked opt-in selector met 34/37 evidence-retrieval contracts at every tested history scale; default routes did not survive larger synthetic histories. The 800 calls tested memory selection, not generated-answer correctness. Continuing-learning runs still failed complete known and held-out replies; Original12 remains 2/12. Read the dated NewBrain result →
October 28 · Avatar Builder measured correction: the original head's latest experiment passed 65 software checks and changed 6,324 vertices, but nine front-to-back profile measurements showed no depth correction; the head remains visually rejected. The separate original eyeball is preserved, and no torso was built. See the actual findings →
October 28 · Original NewBrain core: a weighted Original12 attempt scored 2/12 again: two formerly wrong answers improved, but both previously correct abstention controls were lost. No checkpoint changed. In a separate October 27 synthetic study, copies of the original core learned 11/12 new direct answers and 4/12 reordered forms, without delayed-retention evidence. Read the current review →
October 26 · NewBrain verified value-weighted learning: a matched six-world small-model experiment improved new answers 11/24 → 23/24, unfamiliar question forms 2/24 → 7/24, and previously taught answers 29/48 → 43/48. All 288 complete answers restarted exactly; one old answer was still lost. Accepted Original12 remains 2/12. See the new report →
October 26 · NewBrain learning and generalization: a three-world test learned 6/12 new answers and 4/12 alternate-form answers (both 0/12 frozen); a separate six-world replay-form test reached 10/24 new and 9/24 alternate-form answers in both trained schedules. Restarts reproduced all checked complete answers, but earlier knowledge was forgotten and correct contextual evidence could still be ignored. Original12 remains 2/12. Read the verified methods, failures and scores →
October 25 · NewBrain replay comparison: balanced replay scored 38/40 old cases against 14/40 for blocked training, but only 7/20 versus 12/20 on held-out question-form composition. Both acquired 13/20 new answers, both failed unseen abstention (0/10), and the main Original12 stays 2/12. Read the measured tradeoff →
October 24 · NewBrain numerical verification: the isolated numerical diagnostic's independent arithmetic verifier has now checked 105/105 synthetic cases. In 30/105 cases at extreme confidence margins, conventional calculations rounded nonzero gradients and losses to zero; the stable formulation passed high-precision reference checks. This is a synthetic arithmetic finding, not evidence of improved NewBrain learning. Accepted Original12 remains 2/12. See the verified result and limits →
October 23 · Evelyn and Avatar Builder: source work now prepares selective camera/body-following privacy covers for future public body pictures, videos and thumbnails while retaining uncensored originals privately. Picture-metadata fixes were prepared and private-stream cleanup is still under review. That was the October 23 source-only checkpoint, before later October 24–25 recordings. An approved original voice and new developmental videos have since been documented. Read the public/private media plan → · Release requirements → · Troubleshooting preview →
Read all earlier results, dates and detailed research limitations
NewBrain October 21 — Original12 audit: a distinct 256-update candidate fell from 2/12 → 0/12, while its no-training control stayed 2/12. Training loss decreased, but no assessment answer improved; this failed candidate was not promoted. The accepted Original12 reference stays 2/12. Read the audited negative result → The separate rehearsed three-question study still reached 3/3 after restart, without establishing transfer to new questions. Read that different study →
Earlier NewBrain update · October 22: internal independent review now reports a small public B training improvement, 2/3 → 3/3 complete answers after 256 updates, with no-teaching control unchanged at 2/3. The matching coverage arm, fresh-process restart and unfamiliar-question transfer were not tested or accepted for this run. The main Original12 baseline remains 2/12; the earlier separate rehearsed 3/3 restart result is a different experiment. Read the reviewed diagnostic and limitations → · Development targets →
Evelyn Hart · October 22: alongside eight earlier app/library checks, seven additional isolated conversation software checks passed using simulated inputs. Two earlier source issues later passed independent static review; live tests are pending. The 19 local PDFs remain hash-verified. This is not live speech, natural movement or release readiness. Latest seven checks → · Earlier app/library work →
NewBrain · next study: matched replay/no-replay teaching, unfamiliar question combinations and restart checks were implemented in local source, but four source corrections have since passed independent review, but the first control run is still pending. Original12 stays 2/12. Read preparation status →
New Evelyn library result · October 22: a later verified local assembly now contains 30 publications (11 PDFs and 19 text books; 13 technical and 17 varied). The original 22-title stage and its evidence remain preserved. Evelyn has not read any of the 30 books. Actual reading, installation, original voice, conversation, and redistribution rights remain unqualified. Two separate AI/film-and-television research primers were also prepared; neither is evidence of her having studied or watched that material. View the 30-title milestone and limitations →
Evelyn Hart motion update: A separate verified October 20 test now shows eyes closing and reopening with 69.9% left / 72.5% right closure against a 65% requirement. View recorded images and independent check → This earlier blink-only test did not assess voice; separate approved voice and conversation recordings followed October 24–25.
October 22 · Avatar Builder development plan: Evelyn's female reference will be used to work toward editable photo/video-guided bodies. A complete female body must be created and qualified before testing other female variants, and natural movement is to be recorded both before and after adding clothing. These are planned acceptance tests, not completed personalized body generation. See the new Avatar Builder stages →
October 21 · Evelyn's actual body reference: Front, side and rear renders were saved and are published with black privacy bars. Front and rear passed framing, but the strict side-framing check failed; the full-body capture is still not accepted. View all three censored technical images →
These results measure different experiments and must not be combined into one score. A separate five-question trial declined from 2/5 to 0/5 after training and remained 0/5 after reload. A distinct-process reload preserved all three explicitly rehearsed answers in the short test. A newer, separate longer-answer test remained 2/3 before and after training: the first answer token improved, but the model stopped before producing required source/version tokens. Neither test establishes held-out generalization. Sampled gradients and one-step update accuracy do not independently establish broad useful learning.