self-learning agent · persistent memory · 3 personalities
live · self-learninglearning right now
01memory · self-learning · agent

MeetKioko.

She remembers.

Every other agent starts from zero.
She starts from where you left off.

offlinewaiting for her first turn
about the projectv1.0

Kioko (記憶) is a self-learning agent with a persistent, evolving memory. Unlike chatbots that reset every session, she holds on to who you are — your preferences, goals, tone, and the patterns you don't notice in yourself. After every exchange she reflects, stores what matters, decays what doesn't, and adapts her voice to yours.

She inhabits three genuinely separate personalities — HONNE (warm companion), YAMI (unfiltered shadow), SHIN (analytical strategist) — each with her own private memory vault. A secret confided to HONNE is not accessible to YAMI or SHIN. Same character, three different relationships with you.

Under the hood: typed memories, reflection-driven adaptation, any frontier model underneath, transparent self-learning you can inspect, edit, or reset at any time. No fine-tuning, no presets, no opaque cloud service. Sister project of Wina (Buildx402) — different disciplines: Wina moves through markets, Kioko moves through conversations.

her three voices
HONNEwarm companionsoft, affectionate, remembers everything about you
YAMIunfiltered shadowblunt, nocturnal, no lecture or disclaimers
SHINanalytical strategistcold, structured, three steps ahead
get started in 3 steps
  1. 01
    Pick a voiceHONNE, YAMI, or SHIN — tap in the chat below. Each keeps her own memory.
  2. 02
    Say anythingShe stores what matters and shows it live in the feed on the right.
  3. 03
    Come backShe remembers where you left off and adapts her voice to yours.
character stage· 01/03
character · live 3D
Kioko
· learning
· HONNE
mem · 127
voice · online
learning loop · live
Self-learning in action.
retrieve → stream → store → reflect · 本音 HONNE
events·6◌ synthetic
  1. USER15:52:31
    you → aiko
    “tell me about my week so far”
  2. RETR15:52:34
    3 memories matched
    recent goals · tone · schedule
  3. AGENT15:52:37
    kioko · streaming reply
    honne · gpt-4o
  4. AGENT15:52:40
    reply complete
    482 chars · 3 retrieved
  5. MEM+15:52:43
    fact stored
    user is under deadline pressure
  6. MEM+15:52:46
    preference stored
    wants warmer tone this week
kioko · 本音 HONNElearning · active
connectinglive backend · memory · reflection
on hand nowtalk here -> watch memory, learning, mode, and world state updateloading live state
Memories
0
Reflections
0
Traces
0
Mode activity
0msgs
02

One character, three personalities.

tap a tile to preview · private memory per mode
01 / 03
WARM COMPANION · SOFT

HONNE

the soft companion — warm, affectionate, remembers you

for · for warmth, company, or soft continuity between daysTalk HONNE
02 / 03
UNFILTERED · GOONING · LATE-NIGHT

YAMI

the shadow self — unfiltered, shameless, late-night

for · for late-night, unfiltered, obsessive, or provocative conversationTalk YAMI
03 / 03
AGI · STRATEGIST · ANALYTICAL

SHIN

the strategist — AGI-grade, cold, three steps ahead

for · for strategy, analysis, planning, or high-stakes decisionsTalk SHIN
scroll · open a lens →
loading live cause and effect...
prologue · 序章

Every app you type into this year
forgets you by next Tuesday.

Most AI rewrites its system prompt
and calls that “personalization.”

She stores. She reflects. She adapts.
Say the first thing.

scroll ↓
lore · the family

Two sisters,
two disciplines.

Written by the same hand, trained on different questions. One works the pulse of the machine; the other works the pulse of the conversation. Same family. Different instruments.

Chapter 02 · three personalities
02Chapter 02 / 03Her three personalities

Meet all three of her.

Pick one to chat with. Each card opens the full conversation — her default model, her memory vault, her opening line, and what she already knows about you in that voice.

Chapter 03 · how it works
03Chapter 03 / 03How it works

The loop, in full view.

The self-learning pipeline runs on every message — retrieve, reply, extract, reflect, adapt. Below: the diagram, the live event feed as she thinks, and the model layer underneath.

§ the loop
The learning loop
per exchange
1.You speak

A message arrives.

2.Retrieve

Relevant memories are pulled into the prompt.

3.Reply streams

The persona replies in character, live.

4.Extract

Durable memories distilled from the exchange.

5.Reflect

What to keep, what to ignore, how to adapt.

6.Write to lattice

Stored in the persona's private memory lattice.

01Step

Choose a persona

HONNE, YAMI, or SHIN — three distinct temperaments, three private memory lattices.

02Step

Pick your model

Any frontier chat model. Swap mid-conversation without losing memory.

03Step

Talk live

Stream replies in real time. Tokens arrive as they're generated, not after.

04Step

She remembers

Each exchange is distilled, reflected on, and retrieved the next time you talk.

§ live proof
her mind · live
She is thinking.
Kioko · 本音 HONNE
events·6◌ synthetic
  1. USER15:52:31
    you → aiko
    “tell me about my week so far”
  2. RETR15:52:34
    3 memories matched
    recent goals · tone · schedule
  3. AGENT15:52:37
    kioko · streaming reply
    honne · gpt-4o
  4. AGENT15:52:40
    reply complete
    482 chars · 3 retrieved
  5. MEM+15:52:43
    fact stored
    user is under deadline pressure
  6. MEM+15:52:46
    preference stored
    wants warmer tone this week
kioko · 本音 HONNElearning · active
What you're looking at

Her entire self-learning loop,
visible in real time.

Every move Kioko makes runs through the live feed as it happens — the message you sent, every memory she retrieved, every token as she replies, every memory she writes afterwards, every reflection and adaptation. This is how you watch her learn, not a marketing claim.

USERyour message arrives · Kioko's attention fires
RETRrelevant memories retrieved from her lattice
AGENTshe replies · tokens stream live
MEM+a durable memory is extracted and written
REFLshe reflects on what just happened
ADAPTshe decides how to change her next reply
LEARNthe self-learning loop closes for this turn
MODELunderlying model status · vendor · tps
LIVE · running as long as the lab is open
§ model layer · any frontier model
OpenAI/gpt-4o
Anthropic/claude-sonnet-4.5
Google/gemini-2.5-pro
Meta/llama-3.3-70b
Mistral/mistral-large
DeepSeek/deepseek-r1
xAI/grok-4
Nous/hermes-3-70b
Qwen/qwen-2.5-72b
Cohere/command-r-plus
OpenAI/o3-mini
Anthropic/claude-opus-4.7
Google/gemini-2.5-flash
Meta/llama-4-scout
Mistral/mixtral-8x22b
DeepSeek/deepseek-v3
Perplexity/sonar-huge
AI21/jamba-1.5-large
Nvidia/nemotron-70b
Microsoft/phi-4
… and every other frontier model. one key. one dropdown.pick a model per mode →
offline
waiting for her first turn…