AI Memory¶

AI that learns from you. Pauhu's AI memory learns from every translation, building contextual understanding of your organization's style, terminology preferences, and domain knowledge.

How AI Memory Works¶

graph TB
    A[Translation Request] --> B[AI checks memory]
    B --> C{Context found?}
    C -->|Yes| D[Apply learned patterns]
    C -->|No| E[Standard translation]
    D --> F[Translate with context]
    E --> F
    F --> G[Store in AI memory]
    G --> H[Update semantic model]

What AI Remembers¶

Memory Type	Example	Retention
Terminology preferences	You prefer "tekoäly" over "keinoäly" for "AI"	Permanent
Style patterns	Formal vs. informal address	Per-project
Domain context	Legal documents use specific phrasing	Per-domain
Sentence structure	You prefer active voice	Permanent
Acronym expansions	First mention expands, subsequent uses acronym	Per-document

Contextual Learning¶

from pauhu import Pauhu

client = Pauhu()

# First translation - AI learns your preference
result1 = client.translate(
    text="AI systems must comply with regulations.",
    target="fi"
)
print(result1.text)
# "Tekoälyjärjestelmien on noudatettava säännöksiä."

# Later translation - AI remembers context
result2 = client.translate(
    text="AI transparency is required.",
    target="fi"
)
print(result2.text)
# "Tekoälyn läpinäkyvyys on vaadittua."
# ↑ Consistently uses "tekoäly" (AI learned your preference)

Session Memory¶

Within a session, AI remembers:

# Document context maintained across chunks
doc = client.documents.translate(
    file="eu-ai-act-full-text.pdf",
    target="fi"
)

# AI remembers:
# - Document title and subject
# - Previously translated sections
# - Defined terms from earlier in document
# - Writing style (formal, legal)
# - Acronyms already expanded

Example: Pronoun Consistency¶

# Chunk 1:
"The data controller shall ensure..."
→ "Rekisterinpitäjän on varmistettava..."

# Chunk 2 (same document):
"They must also document..."
→ "Sen on myös dokumentoitava..."
# ↑ AI remembers "they" = "data controller" (singular in Finnish)

Project Memory¶

Per-project, AI maintains:

project = client.projects.create(
    name="EU Legislation Q1 2025",
    domain="10 European Union"
)

# AI learns project-specific patterns:
# - Preferred terminology variants
# - Style guide compliance
# - Reviewer corrections
# - Glossary terms in context

Memory Updates from Corrections¶

# Original translation
result = project.translate(
    text="machine learning algorithm",
    target="fi"
)
print(result.text)
# "koneoppimisalgoritmi"

# User corrects
project.memory.correct(
    source="machine learning algorithm",
    target_was="koneoppimisalgoritmi",
    target_should_be="koneoppimisen algoritmi",
    reason="Organization style guide"
)

# Future translations remember
result2 = project.translate(
    text="Our machine learning algorithms are transparent.",
    target="fi"
)
print(result2.text)
# "Koneoppimisen algoritmimme ovat läpinäkyviä."
# ↑ AI remembered the correction

Semantic Memory Integration¶

AI memory connects to:

Translation Memory (TM)¶

# AI learns from TM matches
result = client.translate(
    text="data protection impact assessment",
    target="fi"
)

# AI checks:
# 1. TM for exact/fuzzy matches
# 2. AI memory for similar phrases
# 3. Context from both sources
# → "tietosuojaa koskeva vaikutustenarviointi" (learned from TM)

Term Base¶

# AI learns term usage patterns
# Term Base says: "AI" → "tekoäly"
# AI Memory learns: "AI systems" → "tekoälyjärjestelmät" (compound)
#                   "AI-powered" → "tekoälypohjainen" (adjective)

Cross-Language Memory¶

AI remembers patterns across language pairs:

# EN → FI translation
result1 = client.translate(
    text="The regulation applies to all AI systems.",
    source="en",
    target="fi"
)

# EN → SV translation (different target)
result2 = client.translate(
    text="The regulation applies to all AI systems.",
    source="en",
    target="sv"
)

# AI applies similar structure:
# FI: "Säännös koskee kaikkia tekoälyjärjestelmiä."
# SV: "Förordningen gäller alla AI-system."
# ↑ Consistent passive voice pattern learned

Privacy and Data Retention¶

What's Stored¶

# Encrypted in AI memory:
# ✓ Terminology preferences
# ✓ Style patterns
# ✓ Structural preferences
# ✓ Domain-specific knowledge

# NOT stored:
# ✗ Source text content
# ✗ Translated content
# ✗ Personal data

Client-Side Encryption¶

AI memory is encrypted client-side:

client = Pauhu(
    client_side_encryption=True  # Default
)

# Your AI memory patterns are encrypted before storage
# Pauhu cannot read your preferences or patterns

Memory Management¶

# View AI memory statistics
stats = client.memory.stats()

print(f"Learned patterns: {stats.pattern_count}")
print(f"Terminology preferences: {stats.term_preferences}")
print(f"Style rules: {stats.style_rules}")
print(f"Memory size: {stats.size_mb} MB")

# Reset AI memory for project
project.memory.reset()

# Export AI memory (for backup)
project.memory.export("./ai-memory-backup.json")

# Import AI memory (restore from backup)
project.memory.import_file("./ai-memory-backup.json")

Compliance¶

"The data subject shall have the right not to be subject to a decision based solely on automated processing..."

AI memory does not make decisions - it assists human translators with learned preferences.

ISO 17100:2015¶

"Translation resources shall include... reference materials and terminology resources"

AI memory augments traditional TM/term base with learned contextual patterns.

Performance Impact¶

AI memory improves over time:

Translations	Quality Improvement	Speed Improvement
0-100	Baseline	Baseline
100-1,000	+5%	+10%
1,000-10,000	+12%	+25%
10,000+	+18%	+35%

Based on BLEU scores and processing time vs. cold start

Getting Started¶

from pauhu import Pauhu

client = Pauhu()

# AI memory is enabled by default
result = client.translate(
    text="Your content here",
    target="fi"
)

# AI learns from every translation
# Check what AI remembered
print(result.metadata.patterns_learned)
# ["term_preference", "style_pattern"]