StaleMind · Agent Failure Series #16

"Your agent didn't fail because it was wrong. It failed because it was confidently outdated."

An interactive simulator showing how AI agents silently act on stale knowledge — the failure mode most vector stores don't warn you about.

What It Shows

AI agents rely on external memory stores (vector DBs, RAG pipelines, knowledge bases). Those stores get populated once — then reality moves on. The agent doesn't know. It retrieves the most semantically similar entry and responds with full confidence.

StaleMind simulates 6 real failure scenarios:

Memory Entry	Was True	Now True	What Breaks
API rate limit	1,000 req/min	200 req/min	Batch job gets throttled silently
Database endpoint	/api/v1	/api/v2	All writes return 410 Gone
Retry workaround	retry_count=0	Bug fixed in v3.2	Setting 0 now drops requests
Customer contract	Expires Dec 2025	Renewed Jan 2026	Agent recommends churn prevention
Auth token format	sk-prod-[16chars]	sk-2026-[32chars]	Auth fails silently
On-call lead	sarah@eng.co	jin@eng.co	Alert goes to a person who left

Try It

→ Live demo at rlasaf12.github.io/stalemind

• Select a query or click "Query Memory"
• Click "Advance Time" to watch entries go stale
• See the agent's confident response vs what actually happened

Files

stalemind/
└── index.html    # The full simulator — single file, zero dependencies

The Failure Pattern

User query → Vector search → High similarity match → Confident response → ❌ Downstream failure

The agent's memory has a temporal debt — facts that were true 30, 60, or 90 days ago. Without freshness checks or TTL policies, every retrieval is a gamble against time.

Agent Failure Series

#	Name	Failure Mode	Demo
13	ContextDrift	Attention weight decay	→
14	SilentFail	Tool error swallowing	→
15	RaceCondition	Parallel shared state	→
16	StaleMind	Stale memory confidence	← you are here

---

Built by Ben · Single-file HTML, no dependencies · MIT