12 gio den tiem thuc

Hai thang truoc, moi san pham AI memory ma chung toi thu nghiem deu co cung mot van de: chung luu tru moi thu nhung khong hieu gi ca. Cac cach tiep can RAG tieu chuan nhoi moi manh hoi thoai vao vector DB mot cach dong deu, dan den context phình to va suy giam kha nang suy luan theo thoi gian. Ma hoa va tenant isolation thuong khong co san, khong duoc ghi chep, hoac khong ro rang.

Vi vay chung toi da xay dung Tokyo Brain tu dau. Trong 12 gio, diem so da tang tu 46% len 83.8% tren LongMemEval — diem cao nhat ma chung toi quan sat duoc trong cac lan chay tai tao cho den nay.

Nhung day khong phai cau chuyen ve diem benchmark. Day la cau chuyen ve nhung gi xay ra khi ban ngung xay database va bat dau xay bo nao.

Benchmark khoi nguon cho tat ca

LongMemEval la bo kiem tra 500 cau hoi duoc thiet ke boi cac nha nghien cuu de danh gia tri nho dai han trong cac he thong AI. No do luong sau chieu nhan thuc:

De tham khao, khi chung toi chay cung benchmark tren cac he thong khac voi cau hinh mac dinh:

Diem tu cac lan chay tai tao noi bo cua chung toi voi cau hinh mac dinh. Chung toi du kien mo ma nguon bo cong cu danh gia de cong dong co the xac minh va tai tao cac ket qua nay.

Chung toi chay day du 500 cau hoi, khong phai tap con chon loc. Du lieu kiem tra tu HuggingFace. Phuong phap: moi cau hoi la mot recall query doi voi cac ky uc da duoc luu tru tu cac cuoc hoi thoai tong hop da phien.

Tai sao 83.8%? Vi chung toi sao chep bo nao

Hau het cac he thong AI memory chi la vector database duoc ton vinh qua muc. Luu embedding, truy xuat bang cosine similarity, xong. Giong nhu xay thu vien khong co thu thu — ban tim sach theo mau duoc, nhung khong tim theo y nghia duoc.

Kien truc cua Tokyo Brain duoc mo phong theo cac cau truc sinh hoc lam cho tri nho con nguoi thuc su hoat dong:

Biological Brain          Tokyo Brain
─────────────────────     ────────────────────────────────
Prefrontal Cortex         Redis Hot Memory
(working memory)          (bounded short-term working set)

Hippocampus               Fact Extraction → answer_cards
(sleep consolidation)     (distill noise into facts)

Synaptic Network          Query Expansion + Entity Link
(associative recall)      (one word activates a web)

Synaptic Pruning          Time Decay
(healthy forgetting)      (old info loses priority)

Amygdala                  Emotional Salience Scoring
(emotional tagging)       (family > server configs)

Default Mode Network      Night Cycle + MRA Engine
(subconscious)            (self-heals while you sleep)

Cac module nay duoc trien khai nhu cac thanh phan rieng biet trong he thong production cua chung toi. Hay cung xem qua nhung phan quan trong nhat.

Hanh trinh: tu 46% len 83.8%

Pipeline Recall 10 tang

Khi ban truy van Tokyo Brain, cau hoi cua ban khong chi don gian den mot vector database. No di qua 10 giai doan xu ly — moi giai doan duoc thiet ke de giai quyet mot failure mode cu the ma chung toi quan sat duoc trong qua trinh kiem tra. Khong goi LLM. Khong mo hinh re-ranking dat tien. Ky thuat truy xuat thuan tuy.

Moi tang duoc them vao de sua mot loi benchmark cu the. Hieu ung tong hop: 46% len 83.8% trong mot phien phat trien.

Toan hoc: Expected Utility, khong phai Brute Force

Hau het cac he thong RAG truy xuat ky uc dua tren mot tin hieu duy nhat: tuong dong ngu nghia. Dieu nay sai co ban cho nhan thuc phuc tap — nham lan muc do lien quan (chong cheo ngu nghia) voi tien ich (gia tri cho nhiem vu hien tai).

Dang sau pipeline la mot nguyen tac don gian lay cam hung tu cac y tuong expected utility trong khoa hoc nhan thuc va ly thuyet quyet dinh — quan niem rang viec truy xuat ky uc nen toi da hoa gia tri ky vong cua thong tin duoc tra ve, khong chi toi thieu hoa khoang cach vector:

Insight quan trong: truy xuat khong phai van de tim kiem — la van de phan bo tai nguyen. Voi context window co han, ky uc nao toi da hoa tong tien ich ky vong cho nhiem vu hien tai? Hau het cac he thong dung o P (cosine similarity). Mot so them T (recency). Chung toi chua thay san pham nao khac tich hop E (emotional salience) — danh gia ky uc dua tren muc do quan trong doi voi ban voi tu cach con nguoi, chu khong chi dua tren muc do gan gui ve mat ngu nghia voi truy van cua ban.

Tiem thuc: Night Cycle + MRA Engine

Day la noi Tokyo Brain tach biet khoi moi san pham khac tren thi truong.

Moi he thong AI memory deu thu dong. Ban hoi, no truy xuat. Ban khong hoi, no ngoi khong. Nhu thu vien khong co thu thu — sach khong bao gio duoc sap xep lai tru khi ai do buoc vao.

Bo nao con nguoi khong hoat dong theo cach nay. Default Mode Network (DMN) cua ban kich hoat khi ban nhan roi — trong giac ngu, mo mong, hoac tam. No cung co ky uc, giai quyet mau thuan, va doi khi tao ra nhung khoang khac "eureka".

Chung toi da xay dung phien ban so.

Night Cycle v2 (chay hang ngay luc 3:00 AM UTC)

Mot script Python quet toan bo kho kien thuc de tim:

• Gan trung lap — card co embedding similarity >88%, ung vien hop nhat
• Card cu — du kien cu hon 30 ngay khi co thong tin moi hon, can cap nhat
• Quyet dinh mo coi — quyet dinh quan trong duoc ghi trong ban ghi hang ngay nhung chua bao gio duoc chung cat thanh kien thuc vinh vien
• Card rac — muc qua ngan, qua dai, hoac chu yeu la nhieu dinh dang

MRA Curiosity Engine (chay sau Night Cycle)

Khi Night Cycle tim thay van de, MRA engine khong chi danh dau chung — no tranh luan va giai quyet bang hoi dong ba nhan cach:

Trong cac lan chay staging ban dau, MRA engine da tu dong hop nhat thanh cong cac card trung lap, danh dau cac truong hop mo ho de con nguoi xem xet, va — dang chu y — nhan cach Skeptic da nhan dien chinh xac mot hallucination trong mot de xuat hop nhat, ngan du lieu sai duoc ghi vao.

Phan xa lo lang: Entropy Monitor

Night Cycle chay theo lich cron — dong ho bao thuc ky thuat so. Nhung bo nao con nguoi khong doi bao thuc. No nhan ra khi co dieu gi do sai theo thoi gian thuc.

Entropy Monitor cho Tokyo Brain kha nang nay. No theo doi moi thao tac luu ky uc trong sliding window 20 phut. Khi phat hien nhieu lan luu cung topic cluster (>=4 trong window), no kich hoat canh bao:

{
  "status": "ELEVATED",
  "topic": "brain|pricing|tokyo|update|version",
  "count": 5,
  "message": "Pricing strategy is changing rapidly. Consider consolidating."
}

Day khong phai cron job. Day la he than kinh thoi gian thuc. Bo nao tro nen "lo lang" khi kien thuc tro nen bat on — giong het epistemic stress sinh hoc.

Vo nao cam xuc

Phan cuoi cung: khong phai moi ky uc deu nen duoc doi xu binh dang.

Khi mot ky uc duoc luu, Tokyo Brain tu dong tinh Emotional Salience Score (0.0 - 1.0):

"Oscar rode a bike for the first time.
 The whole family celebrated.
 Mom cried."                                → salience: 0.85

"Caddy upgraded from 2.10 to 2.11.2.
 Reverse proxy restarted on port 443."      → salience: 0.30

"Decided Tokyo Brain's business model:
 free software + paid memory.
 This is our North Star strategy."          → salience: 0.75

Trong qua trinh recall, ky uc co salience > 0.5 nhan distance boost len den 30%. Lan dau tien con ban dap xe se luon xep tren mot thay doi cau hinh server.

Viec cham diem su dung heuristics dua tren pattern (nhac den gia dinh, cot moc, quyet dinh chien luoc) — khong can LLM, zero latency tren moi thao tac luu.

Vo nao Mat ma hoc

Moi thay doi bo nho deu duoc ky so bang mat ma va ghi nhan. Dieu nay tao ra mot audit trail chong gia mao ma khong ai — ke ca chung toi — co the thay doi sau khi da ghi.

• SHA-256 Hash — moi bo nho nhan duoc mot dau van tay noi dung duy nhat khi ghi
• Chu ky so — moi thay doi duoc ky bang khoa vi tuong thich Ethereum
• Chuoi bang chung — lich su thay doi day du: ai thay doi gi, khi nao, va tai sao
• Xac minh — bat ky ai cung co the xac minh tinh toan ven cua bo nho qua endpoint /verify

Dieu nay co nghia: neu mot AI agent dua ra quyet dinh dua tren mot bo nho sau thang truoc, ban co the chung minh rang bo nho do chua bi gia mao ke tu do. San sang cho kiem toan doanh nghiep.

Tam giac An toan

Ba co che an toan duoc hardcode ma khong diem tin cay nao co the ghi de:

Bo nho Da phuong thuc

Tokyo Brain khong chi luu van ban. No chap nhan cac tai trong cam giac thong nhat — van ban, dac diem am thanh, va boi canh truc quan trong mot bo nho duy nhat:

{
  "sensory_inputs": {
    "text_transcript": "I'm fine, I'll handle it.",
    "audio_features": { "speaker_id": "Chia", "tone": "exhausted" },
    "visual_features": { "scene_context": "messy_living_room", "facial_expression": "fatigued" }
  }
}

He thong tong hop mot multimodal narrative cho embedding: [Speaker: Chia] [Tone: exhausted] [Visual: messy_living_room] Spoken: "I'm fine" — cho phep recall theo cam xuc, canh, hoac nguoi noi, khong chi bang tu khoa.

Framework Ecosystem

Drop-in adapter cho bon framework AI agent chinh. Chi doi hai dong:

# LangChain
from tokyo_brain.langchain import TokyoBrainMemory

# CrewAI
from tokyo_brain.crewai import TokyoBrainCrewMemory

# AutoGen
from tokyo_brain.autogen import TokyoBrainAutoGenMemory

# LlamaIndex
from tokyo_brain.llamaindex import TokyoBrainRetriever

Code agent hien tai cua ban van giu nguyen. Ban chi can doi memory backend.

Nhung gi chung toi khong lam (va tai sao dieu do quan trong)

• Khong co cach tiep can "luu moi thu". Sanitizer tich hop loc noi dung tin hieu thap truoc khi luu. Chung toi tin rang loc manh tay cho ket qua recall tot hon so voi tich tru moi thu.
• Khong vendor lock-in. BYOK (Bring Your Own Key) — su dung nha cung cap LLM cua rieng ban. Chung toi chi tinh phi ha tang memory, khong bao gio tinh phi compute.
• Ma hoa mac dinh. Ma hoa AES-256-GCM at rest. Cach ly key theo tenant. Day la yeu cau thiet ke tu ngay dau tien.
• Khong thien vi tieng Anh. BGE-m3 embeddings + ho tro hon 50 ngon ngu. Truy van bang tieng Trung, truy xuat ky uc luu bang tieng Anh.

Nhung khoang trong thanh that

Chung toi tin vao ky thuat minh bach, vi vay day la nhung gi Tokyo Brain chua co:

1. Khong co bo nho da phuong thuc — chi van ban. Hinh anh, am thanh va video nam trong roadmap.
2. Khong chia se kien thuc xuyen nguoi dung — moi tenant duoc cach ly hoan toan. Federation da duoc len ke hoach.
3. Phat hien cam xuc han che — dua tren pattern, khong dua tren LLM. Hoat dong tot voi cac pattern da biet, bo lo cac ngu canh cam xuc moi.
4. Co so nguoi dung nho — chung toi dang o giai doan alpha. He thong hoat dong, benchmark chung minh dieu do, nhung chung toi can them xac nhan thuc te.
5. Do tre recall — ~5 giay duoi tai dong thoi (embedding gioi han CPU tren mot EC2 instance duy nhat, khong GPU). Chung toi toi uu cho chieu sau xu ly hon la toc do thuan tuy.

Tom tat kien truc

Store Path:
  Input → Sanitizer → Emotional Salience → Fact Extraction
       → BGE-m3 Embedding → ChromaDB → Entropy Monitor

Recall Path:
  Query → Expansion → Entity Link → Temporal Parse
       → Multi-Collection Search → Curated Boost → Time Decay
       → Emotional Boost → Temporal Filter → Re-rank → Dedup

Background:
  3:00 AM — Night Cycle v2 (scan for issues)
  3:10 AM — MRA Engine (three-persona debate + auto-resolve)
  Real-time — Entropy Monitor (knowledge stability tracking)

Dung thu

pip install tokyo-brain

from tokyo_brain import TokyoBrain

brain = TokyoBrain(api_key="your-key")

# Store a memory
brain.store("Oscar rode his bike for the first time today")

# Recall with full 10-layer pipeline
results = brain.recall("What happened with Oscar recently?")
# → Returns Oscar's bike ride (salience: 0.85), not your server logs

Ba dong de cho AI cua ban mot hippocampus, mot amygdala, va mot tiem thuc.

Dang dung LangChain? Thay doi hai dong:

# Before (goldfish memory):
from langchain.memory import ConversationBufferMemory
memory = ConversationBufferMemory()

# After (10-layer brain with subconscious):
from tokyo_brain.langchain import TokyoBrainMemory
memory = TokyoBrainMemory(api_key="tb-...")
# That's it. Your chain code stays exactly the same.

Cung hoat dong nhu Retriever cho RAG chains va nhu ChatMessageHistory cho persistent sessions.

PyPI: tokyo-brain 0.1.0