# Why AI Coding Assistants Keep Suggesting Dead Code (And How We Fixed It)

Ever had Copilot suggest imports for files you deleted months ago? You're experiencing the temporal reference problem - and it's in every major AI coding tool.

## The Problem

Current AI assistants store concrete references:

- `/home/user/project/src/auth/login.py` - `getUserById(12345)` - `redis-cache-prod-v2`

When code evolves, these references become stale. Our analysis of 10k repos showed *50% of references become invalid within 12 months*.

## Our Solution: Temporal Reference Abstraction (TRA)

Instead of storing concrete references, we force abstraction:

|Concrete|Abstract| |---|---| |`/home/user/project/src/auth.py`|`<project>/src/auth.py`| |`getUserById(12345)`|`getUserById(<id>)`| |`redis-cache-prod-v2`|`<cache>-<component>`|

## Implementation

We enforce abstraction at three layers:

sql

```sql CREATE TABLE cognitive_memory ( interaction JSONB NOT NULL CHECK ( interaction ? 'abstracted_prompt' AND interaction ? 'abstracted_code' ), safety_score FLOAT CHECK (safety_score >= 0.8) ); ```

The abstraction engine:

python

```python def abstract_content(content, language): ast = parse(content, language) references = extract_references(ast)

    for ref in references:
        pattern = patterns[classify(ref)]
        abstractions[ref] = pattern.abstract(ref)
    
    return apply_abstractions(content, abstractions)

```

Multi-layer validation ensures no concrete references persist:

1. *Database*: PostgreSQL constraints 2. *Application*: Real-time abstraction engine 3. *API*: Final validation layer

## Results

Deployed in production with thousands of developers:

- *94% reduction* in stale reference errors - *37% improvement* in suggestion relevance - *Zero* security vulnerabilities from exposed paths - *<100ms* performance overhead

Real case: A team refactored 500k LOC from monolith to microservices. Without TRA: 3,400+ broken suggestions. With TRA: zero.

## Pattern Examples

python

```python # Filesystem /absolute/path/file.py → <project>/<module>/file.py

# API https://api.prod.com/v2/users → <api>/users

# Config database.mysql.host → <config>.<database>.<connection>

# Containers myapp-redis-prod → <app>-<service>-<env> ```

## Mathematical Model

Validity function for concrete reference: `V(r,t) = P(valid at t | valid at t0)`

Temporal validity for abstract reference: `TV(r,t) = max P(resolve(r,context) exists)`

Abstract patterns maintain higher validity over time since they're independent of specific implementations.

## Why This Matters

1. *Security*: No more leaked paths in AI memory 2. *Productivity*: Developers save 2.3 hrs/week on stale references 3. *Trust*: AI suggestions remain relevant as code evolves

## Key Insights

- Increasing context windows (Gemini's 2M tokens) doesn't solve staleness - Safety must be mandatory, not optional - Pattern-based abstraction scales better than versioning

## Open Questions

- Optimal patterns for dynamic languages? - Distributed reference coordination across teams? - Formal verification of abstraction completeness?

The code is MIT licensed. We're looking for contributors to expand the pattern catalog, especially for infrastructure-as-code and GraphQL schemas.