Research: long-term evolution of elfmem

Compiled research from May 2026 exploration of how elfmem's memory dynamics compound over years of usage. Drove decisions in ADRs 0001, 0002, 0003.

This research is preserved but is not the active implementation roadmap. For what's actually being shipped, see ROADMAP.md at the project root.

Contents

File	What it covers
closed_form_analysis.md	Six analytical derivations (D1–D6) of structural problems
scoring_proposed_evaluation.md	Empirical validation of plan_memory_scoring.md proposals (lives one level up, applies broadly)
constitutional_evolution.md	Architecture M, Model C, Model D explorations
self_architecting_agent.md	Continuous-parameter framing; hill-climbing simulation
self_critique.md	Honest evaluation of the above against project principles
decisions.md	Decisive synthesis: what to ship, what to defer

Simulation harness

The simulation infrastructure that produced these findings lives at scripts/longitudinal_sim/. It is a permanent fixture for evaluating future scoring changes — not throwaway exploratory code.

Reading order

If you're reading this fresh:

1. Start with closed_form_analysis.md — establishes the problem space mathematically
2. Then scoring_proposed_evaluation.md (parent dir) — what got validated
3. Then constitutional_evolution.md — what didn't earn its place
4. Then self_critique.md — honest re-evaluation
5. Then decisions.md — the verdict

Why this research is preserved

The mechanisms explored here (Architecture M, ego_strength, distribute_n, self-architecting hill-climber) didn't make it into the v0.17 roadmap because the simulations didn't produce decisively better results than the disciplined v0.17 baseline. But the research is preserved because:

• The structural insights (constitutional vs task knowledge as different kinds; signal inflation; asymptotic Bayesian inertness) are correct and may inform future decisions
• The simulation harness is reusable for any future scoring experiment
• The negative results save future contributors from rediscovering them

If you're proposing a new mechanism in this space, start by reading these notes and confirming you have an argument the simulations didn't already explore.