Open-source Engram memory engine, modeled on databases, matches Mem0 on BEAM
phecdaYm · reddit · 2026-10-18
Developer phecdaYm open-sourced Engram, an agent memory engine built on database-style architecture with no LLM or classifier inside.
Design
- Agents store conversations with a short summary and dated facts; the engine keeps the original losslessly compressed and indexes summaries, facts, and chunks via embeddings
- Retrieval picks a layout from the question's shape: time questions get dated facts first, some get full conversations, others summaries plus best passages
- Returned text is marked so recalled memories aren't re-stored as new info
Results
On BEAM 128K split (400 questions, gpt-4.1-mini answering, 1,300-token memory budget): Engram scored 56.8 vs Mem0 55.1 and plain RAG 51.6. Engram was better on time questions and knowing when the answer isn't there; Mem0 led on following user preferences.
A surprise finding: re-storing recalled memories caused a feedback loop that dropped accuracy 6-15 points; marking recalled text fixed it. Author asks for feedback on scalability, evaluating retrieval separately from the answering model, and the Mem0 comparison's validity.
More from coding & agent
- KNOWS benchmark: best web agent fully succeeds in fewer than 3% of complex long-horizon tasks — anmarasovic · 2026-10-18
- Filmmaker ditches Premiere after 25 years by building own editor with Grok in 3-4 days, plans to open source — elonmusk · 2026-10-18
- vLLM and NVIDIA push Kimi K3 decode to 842 tokens/s on GB300 with fused CuTe DSL path — vllm_project · 2026-10-18
- The three ambitions of AI agent memory: personalization, self-improvement, org memory — TheTuringPost · 2026-10-18
- Every Citation Right, Answer Still Wrong 9/9: Code-Checked Conditions Fixed RAG to 5/6 — International_Elk_33 · 2026-10-18
- VFX pro builds entire Houdini scene by voice via Claude, 50-hour render on RTX 5090 — bilawalsidhu · 2026-10-18