一起开一家小咖啡馆
打个招呼，聊做饭，再邀请 Milo 为想象中的咖啡馆起名字。看看他会怎样接话。
A TEACHABLE AI CHILD
A little curious.
A lot to learn from you.
A story, a silly idea, or just a hello. We can start there.
教他一个只有你知道的习俗
从不知道，到学会“月铃礼”，再演示漏记、纠正和回忆。右侧可以看到真实记忆。
0Memories learned
Only what you teach—Current mastery
Changes with time and reinforcement初识Learning stage
Each lesson helps me growMilo AI LEARNER
Curious, transparent, and still learning.
PRESENTATION REQUIREMENT 03
Decisions you can inspect.
The controller manages memory and actions. The language model expresses each selected action.
01ParseStructured input or grounded spans
→02Validate & updateDeterministic memory transaction
→03Retrieve & chooseThis profile only; explicit rules
→04RealizeNatural speech from selected action
→05Log & persistTrace, history and metadata
Live decision trace
Same engine as the chatHow the policy works
priority = 0.5(1 − mastery)
+ 0.3(staleness) + 0.2(relevance)
+ 0.3(staleness) + 0.2(relevance)
Current mastery = stored strength × 2^(−days since reinforcement / stability days), measured at whole-hour resolution. Staleness combines turns since recall with time-based forgetting. Direct questions and autonomous review use the same recall rule. Every fifth question-selection turn explores a new concept. The draw u is a hash, so every decision can be replayed. Bands: low < 0.30 ≤ medium < 0.70 ≤ high.
if a memory has a pending conflict:
clarify()
elif exploration turn (every 5th):
ask_to_learn(quoted mention or new concept)
else:
item = highest priority
u = hash(profile, item, turn)
low: ask_to_learn()
medium: 40% recall, 30% unsure,
30% controlled error
high: 90% recall, 5% unsure,
5% controlled error
# Autonomous correct recall at high mastery:
# practising → ask about use
# connecting → ask about a related concept
# Operators use taught content only:
# omit_step, swap_steps,
# confuse_related, approximate
# Never invent a replacement fact.Teaching starts at 0.35 with a 3-day half-life. Repetition adds 0.10; confirmation adds 0.15; correction adds 0.20 to current mastery (minimum 0.35 after a reminder, maximum 1). Each reinforcement extends the half-life ×1.5, capped at 90 days. Recall alone gives no learning credit. Stages require breadth and reinforced lessons, not elapsed time or chat count. These are illustrative design parameters, not validated estimates of child development.
Topic / keyword graph
Edges link concepts in the same domain that share a keyword or mention each other. Quoted terms (e.g. “kopi”) become curiosity questions.
Content-free interaction log
Metadata is separate from conversation content. No diagnosis is inferred.
PRESENTATION REQUIREMENT 04
Evidence, not promises.
Run the actual controller tests and reproducible synthetic behavioural experiments.
01 · Functional evaluation
96 scripted controller checks, executed in this browser. No live-provider tests.
Run the suite to inspect measured pass counts. No assumed 100% scores.
02 · Personal-baseline change detection
First 20 days calibrate the baseline. Monitoring begins on day 21; a scripted change starts on day 31.
Standardized change score
Lower response rate, longer delay, shorter replies, fewer self-started turns and corrections → higher score.
Cumulative deviation
Sₜ = max(0, Sₜ₋₁ + zₜ − k); signal when Sₜ > h.
03 · Repeated-run robustness
200 stable + 200 sustained-change + 200 three-day-dip sequences, plus a threshold sweep. The default h was fixed in advance.
A single attractive trace is not an evaluation. Run repeated simulations to expose false alarms and misses.
What these results do — and do not — show
IMPLEMENTATION EVIDENCE
A credible teaching loop
Boundary tests, recall, corrections and profile isolation show whether the implemented controller behaves as specified.
SIMULATION EVIDENCE
A visible change signal
False alarms and detection delay quantify the detector on our declared synthetic generator, not on elderly users.
STILL A HYPOTHESIS
Meaningful human benefit
Engagement, loneliness, authenticity and safety need consented user research. They cannot be concluded from this demo.
OPTIONAL BY DESIGN
Connect when you're ready.
The public demo works without an API key. Provider credentials stay on your server.
Connect the backend
Enter your ChildGPT backend, not a model-provider endpoint. GitHub Pages hosts only the static frontend.
Not connected. API configuration is intentionally empty.
Constrained API mode
The model uses our recent conversation to recognize teaching, questions, corrections and confirmations. The controller validates references and new facts before updating memory. It writes natural replies from the selected action and supplied memory. It is instructed to ask about untaught knowledge; generated wording still needs evaluation.
Server-only configuration
Copy .env.example to .env in the source project. All provider fields are shipped empty.
LLM_BASE_URL=
LLM_API_KEY=
LLM_MODEL=
APP_ACCESS_TOKEN=
ALLOWED_ORIGINS=Restart the Node server after editing. A configured provider requires a separate application token (24+ characters). Never put the provider key in a frontend file, Pages secret injection, URL or public repository.
Requests require the Chat Completions-compatible endpoint and JSON output support. API errors fall back to deterministic replies, with the fallback reported in the trace.
Your demo data
Local profiles are logical demo partitions, not authenticated identities. Backend sessions add separate bearer-token access; production accounts are out of scope.
Scope and limitations
This is a software prototype for GP8001 Group 6, not a care service. Use invented examples, not personal, medical or identifying information.
Memory is explicit. A connected model writes natural replies within the selected action; templates remain available when the model is unavailable. Controlled errors follow explicit recall rules; a scripted omission is also available for demonstrations.
Optional invitations work only while this page is open. The backend has a seven-day session lifetime. Consent, age-appropriate field research, clinical validation and production security require separate work.
No analytics, third-party fonts or external scripts are embedded in this app. Public hosting has its own server logs.