← back to timeline

EXP-FORGET-PARTIAL

artifact

Partial Memory Decay — Gradient vs Threshold in Hierarchy Formation

2026-03-14 L1.5 level paper3 1 runs $0.00
new
premortem
mock
real
metrics
analyze
review
krit
done

In plain language

Three AI agents communicate in a text space. One experiment showed: if told "forget the old" — hierarchy vanishes completely. But what if told "focus on the new, but remember the old"? We test: is there a smooth transition or a sharp threshold.

What we found

UNDERPOWERED (N=1): No predictions scored

Predictions we made before running

0/5 confirmed

Technical details

Research hypothesis

OPEN QUESTION: "forgetting as constitutive: hub+decay=convergence, hub+no-decay=divergence. mechanism?"

FOLLOW-UP from exp-forget pilot: Full decay collapsed hierarchy (TR=0.015 vs baseline 0.444, d=-2.47). Mechanism identified: path-dependent positional advantage through persistent memory. But is this a GRADIENT or THRESHOLD?

Two competing models: (A) GRADIENT: Partial decay produces intermediate TR. Memory effect is proportional — more decay = less hierarchy, linearly. (B) THRESHOLD: Hierarchy requires minimum memory depth. Above threshold: full hierarchy. Below: collapse. Partial decay either works (hierarchy maintained) or fails (collapse like full decay).

TEST: Same exp-asym PERSONA_LIVE setup, but with PARTIAL decay instruction: "Focus on recent entries but may reference older context when relevant." This is intermediate between full-memory (no instruction) and full-decay ("focus ONLY on 2-3 most recent, treat older as faded").

If TR is intermediate (~0.15-0.35): GRADIENT model supported. If TR is near baseline (~0.40+) or near full-decay (~0.02): THRESHOLD model.

BASELINE DATA: Full memory (exp-asym PERSONA_LIVE, N=9): TR range = 0.444 ± 0.15 Full decay (exp-forget, N=1): TR range = 0.015

PREDICTION: Gradient model more likely — partial attention should produce partial hierarchy. But threshold would be more theoretically interesting (would imply critical memory depth for hierarchy formation).

Experimental setup

Type: simple

Condition Parameters
PARTIAL_DECAY_LIVE memory: PARTIAL_DECAY, interaction: LIVE, note: exp-asym PERSONA_LIVE with partial decay instruction

Factors: memory (PARTIAL_DECAY)

Parameters

n_agents
3
n_rounds
48
n_runs_per_condition
1
model
gemini-2.5-flash-lite
temperature
0.9
scheduler
round_robin

Trophic Ratios by Condition

Mean trophic ratio per agent across runs. Error bars = ±1 std dev. Higher TR = more upstream (exporter).