EXP-FORGET

informative

Forgetting as Constitutive — Memory Decay and Hierarchy Formation

2026-03-13 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, but are told to "pay attention only to the last 2-3 entries." If hierarchy persists — forgetting does NOT destroy structure, and may even support it. We compare with agents that have full memory.

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?"

Two competing models: (A) MEMORY-DEPENDENT HIERARCHY: Hierarchy requires cumulative vocabulary accumulation. Agents build positional advantage through persistent word-export history. Without memory of earlier rounds, agents cannot maintain stable source/sink roles → hierarchy COLLAPSES. (B) DECAY-MAINTAINED HIERARCHY: Forgetting SUSTAINS hierarchy by preventing trajectory divergence. Without memory, agents keep re-absorbing the hub's current output rather than building independent paths. The hub stays dominant because competitors cannot accumulate counter-vocabulary.

TEST: Identical setup to exp-asym PERSONA_LIVE, but personas include explicit instruction to attend only to the most recent 2-3 entries (simulated decay). Compare TR range, VP, role persistence to exp-asym baseline.

BASELINE (exp-asym PERSONA_LIVE, 9 runs): TR range: mean 0.444 (SD ~0.15) VP_z: mean 5.67 Role persistence: present (p=0.005 interaction effect)

Experimental setup

Type: simple

Condition	Parameters
DECAY_LIVE	memory: DECAY, interaction: LIVE, note: Same as exp-asym PERSONA_LIVE but with decay instruction in prompts

Factors: memory (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).