MemBukkit: An Open-Source Long-Term Memory Layer That Beats Fancy Setups With a Boring Flat Index

AOZakari · reddit · 2026-08-28

An XAI PhD student, fed up with memory solutions claiming SOTA on questionable evals and collapsing on real projects, open-sourced MemBukkit (Apache 2.0, runs locally, works with open models).

Core insight: semantically similar ≠ useful for answering. Ask "when is my family free next week" and semantic search happily returns potato-salad trivia while the answer hides in an unrelated shift-change conversation. The common fix — throwing a reasoning model at the whole memory — is an expensive way of admitting retrieval sucks. MemBukkit flips this: the stored side stays deliberately boring (dated facts + original source, flat index, optional buckets, no LLM-authored graph to rebuild), while retrieval is trained for the task, with the access policy chosen by whether retrieved context actually helps the reader answer.

Results: 88.8% on LongMemEval-S with Gemma 4 26B as the open-weight reader; 92.6% with GPT-5.4 under the official judge; 87.5% zero-shot on LoCoMo, and the flat-index idea carries to multi-hop RAG. A fun ablation: plain cosine beats some fancy reranking setups. The author is building a company around it but keeps the core open-source.

Original post →

More from coding & agent

coding & agent channel →