# Your AI stack is carrying Cache Debt.

> Cache Debt is the hidden cost created by stale AI context, bloated prompts, outdated embeddings, old model cache, and forgotten agent artifacts.

Canonical page: https://www.cacheddebt.com/

Find stale context, bloated prompts, forgotten embeddings, old agent artifacts, and model cache waste before they slow everything down.

- [Read the definition](https://www.cacheddebt.com/glossary/cache-debt/)
- [View the checklist](https://www.cacheddebt.com/checklist/)

## Measured, not asserted

-   [**920 MB**of browser builds no test used](https://www.cacheddebt.com/score/#worked-example)
-   [**39 MB**of plaintext agent transcripts, scored Act now](https://www.cacheddebt.com/score/#worked-example)
-   [**9**tools with documented paths and commands](https://www.cacheddebt.com/checklist/)
-   [**15**primary sources, quotes checked against the live page](https://www.cacheddebt.com/sources/)

Sizes measured with the read-only audit script on the machine that built this site. One machine, not a benchmark.

## AI systems remember more than teams maintain.

We define Cache Debt as the hidden cost created when AI systems keep stale context, bloated prompts, outdated embeddings, old model cache, unused artifacts, and forgotten agent state.

It accumulates quietly when retained material grows faster than a team can review, refresh, or remove it.

[Read the complete definition](https://www.cacheddebt.com/glossary/cache-debt/)

## Old state becomes active friction.

Cache Debt is not only a storage problem. It changes the cost, speed, quality, and explainability of AI work.

1.  01
    
    **It raises operating cost.**
    
    Redundant files, embeddings, logs, and model caches consume storage and compute without adding current value.
    
2.  02
    
    **It weakens output quality.**
    
    Stale context and old prompt versions can steer models toward answers that no longer match the system.
    
3.  03
    
    **It hides the cause of failures.**
    
    When retained state has no owner or history, teams spend longer separating current behavior from forgotten residue.
    

## Every retained layer can drift.

Look beyond one cache folder. Debt forms wherever AI systems preserve information without a review cycle.

### Context layers

Saved instructions and conversation state keep accumulating after their useful life ends.

### Prompt systems

Old prompt versions remain active, duplicated, or detached from the behavior teams expect.

### Vector stores

Embeddings drift away from their source material without a refresh or deletion policy.

### Agent workspaces

Logs, tool results, generated files, and temporary state become permanent by accident.

## Make retention an explicit decision.

Four steps turn retained state from an accident into a decision someone owns.

1.  1.0 / Step
    
    ### Inventory
    
    Find what the system retains, where it lives, and which workflow still depends on it.
    
2.  2.0 / Step
    
    ### Classify
    
    Separate active material from stale, duplicated, orphaned, or unknown state.
    
3.  3.0 / Step
    
    ### Refresh or remove
    
    Update valuable state and safely delete material that no longer earns its place.
    
4.  4.0 / Step
    
    ### Set retention rules
    
    Give every cache, prompt, embedding, and artifact an owner and review interval.
    

## The work crosses team boundaries.

AI engineers, platform teams, developer-experience teams, security teams, and technical leaders each see a different part of the same maintenance problem.

-   01**AI engineers**
    
    Prompt versions, context windows, and retrieval quality.
    
-   02**Platform teams**
    
    Model stores, vector databases, and shared caches.
    
-   03**Developer experience**
    
    Local caches, agent workspaces, and build output.
    
-   04**Security teams**
    
    Retained transcripts, credentials in context, and log retention.
    
-   05**Technical leaders**
    
    Ownership, review cadence, and the cost of retained state.
