Learning
Change with experience
Small language, plasticity and value-learning components update internal numerical state from controlled lessons and outcomes.
Kira Labs model research · active R&D
NewBrain.
Kira Labs’ first model.
NewBrain is Kira Labs’ first internally developed experimental trainable cognitive model. It is being built as a reusable cognitive architecture for synthetic people: learning, memory, plasticity, value formation, sensory processing and eventually body feedback, while keeping each individual’s identity and history separate.
What kind of model?
A cognitive architecture,
not one giant network.
NewBrain combines multiple trainable and stateful subsystems rather than trying to reproduce a conventional foundation model. Python currently coordinates much of the research implementation while learned numeric state, memory and controlled components are measured separately.
Memory
Keep histories distinct
Episodic-memory research separates stored experiences, learned weights and working context, including a matched episodic-ON / episodic-OFF design for Maya.
Reusable core
One architecture, different people
The goal is a shared cognitive core that can eventually support Maya, Kira, Aster and other synthetic people without sharing their personal identity, memories or relationships.
Measured evidence · October 5, 2026
Promising results.
Narrow claims.
Kira Labs keeps engineering checks, scientific results and future goals separate. Passing infrastructure tests does not automatically become a claim of intelligence.
Repeatable small-task retention: across five repetitions of the same exposed command-learning task, the rehearsal branch retained 216/216 earlier examples and 162/162 new examples in every run. Without review, earlier-task results ranged from 78–105 correct out of 216.
Controlled hearing foundation: 24 exposed pure-Python hearing engineering fixtures passed with zero failures, errors or skips. They validate causal input handling and learning-test machinery; they do not establish real speech, music or open-world sound understanding.
Additional research components: plasticity, value learning, episodic-memory controls, vision experiments, audiovisual benchmark design and restart/state-transfer work are being developed and tested separately so gains can be attributed to the component that produced them.
Measured
Retention on one constrained task
Repeatable across five runs, with the exposure and control conditions preserved in the repository.
Engineering
Hearing test foundation
Twenty-four engineering checks passed; scientific sound-learning and real-audio understanding remain ahead.
Still unfinished
General conversation
NewBrain cannot yet be presented as a general conversational model comparable to Qwen or ChatGPT.
Evidence boundary: the measured retention result uses the same exposed constrained task, not fresh general-language concepts. NewBrain has not demonstrated broad general knowledge, unrestricted conversation, robust vision, real speech/music understanding, a complete body, or a fair overall win against Qwen.
Developmental experiment
Maya tests what
experience adds.
Maya is the planned episodic-memory learner. Her matched control receives the same architecture, lessons, learning rules, weight updates and working-context allowance while conventional episodic storage/retrieval is disabled.
The curriculum is intended to grow gradually through ordinary fictional family, school, language, sensory and problem-solving experiences rather than loading an entire biography at once. The point is to measure whether stored experiences improve later recall, corrections, relationships and decisions.
The developmental target includes age-five-inspired skill tests in communication, story understanding, counting, time/order, turn-taking and pretend play. Those are engineering benchmarks, not a claim that Maya has a human developmental age.
Current Maya boundary
Not born by a benchmark
No full trained-backend conversation, authenticated first taught memory or demonstrated five-year-old equivalence has been established. Advancement is meant to follow measured skills and restart retention, not a calendar.
What comes next
Integrate the pieces
without hiding the source.
The next major step is not a larger claim. It is getting learning, recall, language, senses and eventually a body to work together under matched, reviewable tests.
Dialogue
Free-form text input, relevant generated replies and learned recall on unseen questions.
Restart memory
Close the process, reopen it, and verify that learned state and relevant memories survive correctly.
Hearing & vision
Learn from raw or explicitly attributed sensory inputs, generalize to new examples and keep up with continuous media.
Embodiment
Connect the same cognitive core to a tested Kira World body with movement, sensory feedback and measurable physiology-like simulation.
Primary public evidence
Read the source records.
Five-run retention result ↗ · 24 controlled-hearing engineering checks ↗ · Maya / matched-control public organization ↗.
NewBrain is experimental research. “Kira Labs’ first model” means the first internally developed trainable cognitive model from Kira Labs; it does not mean the first model ever used by Kira Labs, and it does not imply parity with established foundation models.