The Use Case

International engineering teams often have a language problem. The Japanese team stores lessons in Japanese. The Korean team writes in Korean. The shared AI memory is siloed — not because the Brain doesn't know these languages, but because keyword search can't bridge language boundaries. Until now.

How It Works — Without Embeddings

Semantic search with embeddings can find cross-lingual matches, but it requires an OpenAI API key, costs money per lookup, and adds latency. Most Brain users are on the free tier — no API key, no embeddings.

Our approach: a curated technical term synonym map, built directly into the tokenizer. When any document is indexed or any query is tokenized, every recognized technical term gets expanded to its equivalents in all supported languages:

Token: "deploy"
Expanded to: デプロイ (JA), 部署 (ZH), 배포 (KO), نشر (AR), פריסה (HE)
+ bigrams of all CJK variants added to the token stream

Token: デプロイ
Expanded to: "deploy", "deployment"
+ romaji "depuroi" added from katakana converter

The expansion happens at tokenize time — both when indexing (documents get synonym tokens) and when searching (queries get synonym tokens). This creates a shared token space across all 6 language pairs.

The Synonym Map — 130+ Terms

The map covers the technical vocabulary that actually appears in developer Brain entries:

Concept	JA	ZH	KO	AR	HE
deploy	デプロイ	部署	배포	نشر	פריסה
container	コンテナ	容器	컨테이너	حاوية	מיכל
server	サーバー	服务器	서버	سيرفر	שרת
error	エラー	错误	오류	خطأ	שגיאה
auth	認証	认证	인증	مصادقة	אימות
monitor	モニター	监控	모니터링	مراقبة	ניטור

Plus 100+ more: cache, database, build, test, install, log, port, cluster, debug, migration, and more.

Zero-Embedding, Zero-Cost

The entire cross-language lookup is a Map.get() call — O(1), no API, no network, no cost.

Traditional semantic search: 200–400ms latency, $0.0004/1000 tokens
Cross-lingual synonym lookup: instant latency, $0

This runs on the free tier, on every smart_recall call, in every Brain session.

Real-World Example

Team setup: Korean backend team stores lessons in Korean. English-speaking DevOps engineers query the Brain in English.

# Korean engineer stores a lesson
learn_from_attempts:
  topic: "deploy:k8s"
  outcome: "success"
  whatWorked: "배포 실패 원인: 포트 3000이 방화벽에 의해 차단됨. 포트를 열어 해결"

# English DevOps searches the next day
smart_recall("deployment failure port blocked")
→ ✅ Returns the Korean lesson, ranked first

# Symmetric: English lesson found by Korean query
learn_from_attempts:
  topic: "fix:redis"
  outcome: "success"
  whatWorked: "Redis connection timeout fixed: set timeout to 5000ms in config"

smart_recall("레디스 연결 오류")  # Korean: "Redis connection error"
→ ✅ Returns the English lesson

Supported Languages

The synonym graph now covers 6 language families: Japanese (hiragana + katakana), Chinese (simplified), Korean (hangul), Arabic (MSA technical vocabulary), Hebrew, and English. All language pairs are bidirectional.

Cross-language search activates automatically — no configuration, no flag, no API key. If you store in Japanese and recall in English, it just works.

Upgrade

npx @cachly-dev/mcp-server@latest autopilot

Store in Japanese, Recall in English — Cross-Language AI Memory