EXP-MEMORY-SWAP

… pending

Memory swap — is between-group separation memory-stored or agent-internal?

2026-04-17 L5 level paper3 5 runs $0.08

new

premortem

mock

real

metrics

analyze

review

krit

done

In plain language

Researchers are investigating whether an agent's understanding of other groups is stored in a shared memory or is an internal trait of the agent itself. They plan to test this by swapping the memory contents between two groups of agents and observing if their interactions change significantly. The results are currently pending.

▶ Technical details

Research hypothesis

Persona-ablation showed between-group distance persists post-ablation (d=0.79, close to PERSONA 0.84). Candidate mechanism: memory path-dependence — each group's SharedMemory contains its own developed vocabulary, so agents continue to echo group-specific words even without persona prompts. Test: keep PERSONA active but SWAP memories at round 40. Agents suddenly see "their group's memory" = other group's content. - If between-group collapses toward MINIMAL (0.64) → memory-stored - If between-group persists near PERSONA (0.84) → agent-internal (something beyond memory, e.g., persona-induced persistent bias)

Experimental setup

Type: simple

Condition	Parameters
MEMORY_SWAP_40	persona: PERSONA, coupling: COUPLED, n_runs: 5

Factors: coupling (COUPLED)

Parameters

n_agents: 6
n_rounds: 80
n_runs_per_condition: 5
model: google/gemini-2.5-flash-lite
temperature: 0.7
scheduler: round_robin

memory-swap emergence-mechanism diagnostic ladder-1a-prime-3