LLM AGENTS · 7 EXPERIMENTAL CONDITIONS

CARM++

Memory-aware LLM agents — testing how much conversational and task memory actually changes agent behavior, across seven controlled conditions.

LLM AGENTSMEMORY-AUGMENTED AI7 CONDITIONSIN PROGRESS

Paper — not yet written upGitHub — privateDownloads — none yet

01 General Idea

LLM agents that carry memory across a conversation or task face a tradeoff: too little memory and they forget useful context turn to turn; too much and they drown in irrelevant history, burning tokens and attention on things that don't matter anymore. CARM++ is an experiment in figuring out how much that memory setup actually matters in practice — running the same agent across seven controlled memory conditions (from no memory at all, to raw transcript, to more selective retrieval-style setups) and comparing how behavior and task performance actually shift.

The goal isn't to ship one "best" memory architecture — it's to get an honest, apples-to-apples read on which memory choices help, which are dead weight, and which actively hurt, before committing to a design.

02 Status — In Production

ACTIVELY IN PROGRESSThis one's still being built and run — agent loop, the 7 memory conditions, and the evaluation harness are in active development, not yet at the paper/write-up stage. No public repo, results table, or manuscript exists yet, so this page intentionally stays high-level instead of presenting placeholder numbers as if they were real findings. Once there's a working results log, this page gets the full breakdown — architecture, per-condition comparison, failure cases, the works.