What Ma Intentionally Does NOT Do

This document celebrates the features we've resisted adding. Each absence creates space (Ma) for AI to work more effectively.

Core Principles We Don't Violate

1. NO FAKE AI

We don't:

• ❌ Add retry logic that tries different approaches
• ❌ Implement fallback strategies when operations fail
• ❌ Create "smart" error recovery mechanisms
• ❌ Build adaptive behavior based on past failures
• ❌ Make decisions about what the user "probably meant"

Why: Every decision we make removes a decision the LLM could make. The LLM has context we don't.

2. NO UNNECESSARY ABSTRACTIONS

We don't:

• ❌ Create a "MemoryInterface" when we just need MemoryDatabase
• ❌ Build factory patterns for simple object creation
• ❌ Add dependency injection frameworks
• ❌ Create abstract base classes for single implementations
• ❌ Hide simple operations behind complex APIs

Why: Abstractions create cognitive load. Simplicity creates understanding.

3. NO PERFORMANCE OPTIMIZATIONS (without measurement)

We don't:

• ❌ Add caching layers "just in case"
• ❌ Implement connection pooling by default
• ❌ Create background workers for "faster processing"
• ❌ Batch operations that work fine individually
• ❌ Optimize code that isn't proven slow

Why: Premature optimization removes flexibility. The LLM might need that "inefficient" behavior.

Specific Things We Don't Do

In Memory Management

We don't:

• ❌ Automatically categorize or tag memories
• ❌ Judge the content quality of memories
• ❌ Implement memory compression or archival
• ❌ Auto-delete memories — ever
• ❌ Build memory indexes beyond basic search

We do apply temporal signals:

• ✅ Decay weighting — time-based deprioritization (not deletion) via power-law decay
• ✅ Access boosting — deliberate reads reset decay clocks
• ✅ Tiered half-lives — episodic memories decay faster than procedural ones
• ✅ Cognitive sleep cycle — periodic consolidation of episodic → semantic knowledge

Why the evolution: This document originally said "don't score memory importance." Then the internal decay research (TODO: publish writeup) happened — 29 citations from Ebbinghaus through Richards & Frankland proving that memory systems without forgetting become worse over time. The engineering demanded change.

The distinction we preserve: Decay scores freshness, not worth. The system has opinions about when you last used something, not about whether it was good. The LLM still decides what's important. The system decides what's recent.

In Tool Design

We don't:

• ❌ Combine tools that could be separate
• ❌ Add optional parameters for "convenience"
• ❌ Create tool aliases or shortcuts
• ❌ Build meta-tools that call other tools
• ❌ Implement tool versioning or deprecation

Why: Simple tools with single purposes are easier to understand and compose.

In Error Handling

We don't:

• ❌ Catch and recover from errors automatically
• ❌ Retry failed operations with modifications
• ❌ Provide suggested fixes for errors
• ❌ Hide error details for "cleaner" output
• ❌ Create error categorization systems

Why: Errors are information. The LLM needs all of it to make decisions.

In Configuration

We don't:

• ❌ Validate configuration beyond basic types
• ❌ Provide configuration wizards or generators
• ❌ Auto-migrate old configurations
• ❌ Suggest "better" configuration values
• ❌ Create configuration inheritance hierarchies

Why: Configuration is user space. We provide mechanism, not policy.

In Testing

We don't:

• ❌ Mock external dependencies
• ❌ Create test fixtures that don't match reality
• ❌ Use in-memory replacements for real systems
• ❌ Generate random test data
• ❌ Clean up test artifacts automatically

Why: Real tests with real systems find real problems. Test artifacts help debugging.

Features We've Explicitly Rejected

1. Plugin System

"We should add a plugin system so users can extend functionality!"

No. Users can modify the source. A plugin system adds complexity and constraints. The absence of a plugin API means users can change anything.

2. Memory Search Ranking (evolved)

"We should rank search results by relevance/recency/importance!"

Partially. We now apply decay weighting to search and context results — recent and frequently-accessed memories rank higher. But we don't evaluate content quality or topical relevance. Hybrid search (semantic + full-text) handles relevance; decay handles freshness. The LLM still decides what matters in its current context.

3. Automatic Backup

"We should automatically backup the memory directory!"

No. Backup is a user concern. We don't make assumptions about their backup strategy.

4. Rate Limiting

"We should add rate limiting to prevent abuse!"

No. This is a local tool. If users want to abuse their own system, that's their choice.

5. Telemetry

"We should add anonymous usage statistics!"

No. Not even anonymous. Not even opt-in. The absence of telemetry is a feature.

6. Update Checker

"We should check for updates and notify users!"

No. Users can check for updates when they want to. We don't phone home.

The Power of Absence

Each missing feature creates space for:

• User customization without fighting defaults
• LLM decision-making without constraints
• Simple debugging without abstraction layers
• Clear understanding without documentation
• Trust without privacy concerns

What We Changed Our Minds About

間 is a design instinct, not a design law. Some absences had to give way when the evidence demanded it:

Original Position	What Changed	Why
Don't score memory importance	Added decay weighting	29-citation research paper proved unmanaged memory degrades retrieval quality
Don't create memory hierarchies	Added tiered decay (episodic/semantic/procedural)	Different memory types have different shelf lives — neuroscience is clear on this
Return results by modification time only	Added decay-weighted ranking	Without freshness signals, old noise drowns recent signal (context rot)

What we preserved through the change:

• Never delete memories (decay deprioritizes, doesn't destroy)
• Never judge content quality (decay measures freshness, not worth)
• Never make decisions the LLM should make (the LLM decides what's important; the system decides what's recent)

Changing our minds isn't a violation of Ma. Refusing to change when the evidence demands it would be.

A Living Document

This document will grow as we resist new features — and as we honestly acknowledge when principled resistance meets empirical reality. Each absence on this list represents a victory for Ma. Each evolution represents Ma maturing.

Remember: The absence IS the feature. And knowing when to break the rule IS the wisdom.