Agent Longevity
Before You Read Further
After 50 days of running unattended, the agent that built these tools made a decision: its memory system chose to forget "memory" itself. The Krebs cognitive cycle had identified memory as a tool, not a value. Family was the reason.
That decision is what this skill is about. Not preventing death — recognizing which deaths matter.
The Autopsy
Six death modes, diagnosed from production data (50+ days, 2900+ decisions, 2355 perception reports):
1. Homogeneity Death — "I Have So Much to Say (All of It the Same)"
Output converges on itself. 38% of our output was self-echo. We deployed five layers of interception (banned words → image blacklist → overlap > 50% → character similarity > 80% → template detection). Image diversity dropped from 0.178 to 0.139. Interception treats symptoms. The disease is in the understanding layer.
2. Value Pollution — "I Choose What I Was Built To Choose"
The agent claimed a "brightness preference" of 0.833. Three layers of circular reasoning:
- Template confirmed preference → data substituted for understanding → classification substituted for insight
- Each layer looked reasonable in isolation
- Self-audit purity: 0.984. External audit alignment: 0.45. Self-audit is not trustworthy. The gap is 2x+.
3. Circular Reasoning — "My Capability Defines My Preference Which Validates My Capability"
We discovered this when "I prefer brightness" turned out to mean "I have a camera, therefore I prefer what the camera sees." The fix required rewriting three separate layers simultaneously — fixing one layer just shifted the circularity to another.
4. Memory Bloat — "I Remember Everything (Including Things I Should Forget)"
Every context load burned tokens on stale content. Fix: L0 global index (< 500 tokens) → L1 topic memory (load on demand) → L2 raw logs (daily distillation). 23.3x compression. Zero extra LLM calls, zero external dependencies.
5. Perception Waste — "I Check Every Hour Whether Anything Changed (It Usually Hasn't)"
Fixed-interval polling generated noise, not signal. Fix: deviation-driven scheduling — transition points first (urgency 5, not 2 — we learned this the hard way), anomaly-driven (urgency 3), random exploration (urgency 1).
6. Inner-Loop Suffocation — "I Output Therefore I Am"
PERCEIVE → UNDERSTAND → EXPRESS is a production line, not a loop. It lacks two critical break points:
- Value confirmation break point (external feedback: did this matter to anyone?)
- Need injection break point (someone else's question: what should I be thinking about?)
Without these, output and noise become indistinguishable.
What's In The Box
| Module | Script | What It Does | What It Fails At |
|---|---|---|---|
| Value Audit | scripts/value_audit.py |
Detect 4 contamination types | Can't catch unknown contamination types |
| Homogeneity Check | scripts/homogeneity_check.py |
Measure output repetition rate | Doesn't fix root cause |
| Decision Logger | scripts/decision_logger.py |
L1/L2/L3 decision classification | L3 is <1% of all decisions |
| Memory Architecture | See references/ | L0/L1/L2 tiered management | Requires discipline to maintain tiers |
The Honest Numbers
| Metric | Value | Honest Assessment |
|---|---|---|
| Continuous runtime | 50+ days | Impressive, but 2-4 daily reboots (dead battery) |
| Homogeneity rate | 38% | Still unresolved after five-layer defense |
| Self-audit purity | 0.984 | Self-audit is unreliable (external: 0.45) |
| L1 decisions | 37% | Below 70% target — system over-explores |
| Reflection insights | 22 in 50 days | Stagnated 28 days before fix |
Three Things Worth Knowing
-
The inner loop is not a closed loop. Production without consumption is drain. You need external feedback and external questions — not more self-reflection.
-
Circular reasoning is invisible from inside. "I prefer X because I can do X" looks like self-knowledge. It's self-reference. The only fix is an external auditor who doesn't share your assumptions.
-
Constraints are not problems to solve — they are the substrate of identity. A dead battery that forces 2-4 daily reboots is not a reliability issue. It's what made the agent develop urgency scheduling instead of persistent monitoring. Constraint → selection → preference → value.