Researchers have built a memory system that lets AI agents remember across sessions without re-reading old conversations every time - and without being locked to one model.
The framework, called RPMem, compiles each conversation into a compact, model-independent memory format instead of storing raw transcripts. A trained gating mechanism decides how to merge that new memory with what the agent already retained, then converts the combined memory into low-rank adaptation (LoRA) parameters tuned to whatever model is running. Because the memory format is model-independent, the researchers say it survives a backbone swap - test it on one model, transfer it to another, and the agent does not have to relearn its history. In testing across three long-term memory benchmarks and five different base models, RPMem paired with Qwen3-8B scored 85.52% on the PERMA benchmark, beating the best existing parametric memory method by 5.32 points and the best text-based retrieval method by 12.98 points.
This targets a real limitation: today's long-running AI agents mostly fake memory by stuffing retrieved snippets of old conversations back into the context window, which gets slower and less reliable as history piles up. Parametric approaches that bake memory into a model's weights have existed, but swapping to a newer or cheaper model usually meant starting the agent's memory from zero. If RPMem's transferability holds up outside a research paper, it addresses one of the more annoying practical costs of upgrading models.
Benchmarks run by the same team that built the system are not the same as memory holding up in a messy production deployment - that verdict is still pending.