supermemory

None of the 11 agent memory systems in a new paper beat plain RAG over the raw chat history on LongMemEval-s. The paper is Memory Tax (arXiv 2610.34073). The baseline is basically the boring thing: chunk the raw history, hybrid search, a reranker, and let the model ask one follow-up query. With Qwen3.6-27B that scored 81.4. The best memory system, A-Mem, got 79.4. It's not cheaper either. On their GPT-5.6 Luna run, A-Mem spent 117 hours building memory offline. RAG spent about 7 minutes indexing and scored 90.0 to A-Mem's 84.0. The useful part is where the loss comes from. Every summarize or extract step is a chance to drop the fact you'll need later. In OpenViking, 38% of sampled raw turns had no memory unit at all. The rest is retrieval: the fact survives, but it gets split up or rephrased and the query stops matching it. Their fix is small. Keep the raw log, retrieve from it alongside memory, and rerank both. It helped on all 3 systems they tested it on. Fair caveat: the authors call LongMemEval-s retrieval-friendly, so memory has less room to help there. But if you're building agent memory, put RAG over the raw history in your eval table. If you can't beat it, the memory isn't adding anything yet. Paper: arxiv.org/abs/2610.34073

with one command, now your agent can use supermemory and fully run it use it - add memories - search them - forget what's out of date run it - create API keys scoped to one namespace - invite teammates - see usage and logs - connect tools like Notion npx supermemory

Introducing the supermemory memory graph every document becomes atomic facts, and dreaming keeps building the graph in the background - extracting facts - linking related memories - resolving updates, the old version stays as history get one for yourself in the console ↓