Mímir swaps RAG for brain-inspired memory

The strongest idea here is not the neuroscience branding, it is the refusal to treat agent memory like a static database. If the numbers are reproducible, this looks more credible for long-horizon assistants than plain RAG.

• –Hybrid keyword, semantic, and date retrieval is the practical backbone; exact names, timestamps, and recency still matter.
• –Reconsolidation, retrieval-induced forgetting, and Zeigarnik-style tension change behavior over time, not just recall quality.
• –The LongMemEval and MSC claims are encouraging, but they still need public scripts and ablations.
• –The standalone fallback path lowers adoption friction and lets teams start small.
• –The biggest risk is complexity creep: 21 mechanisms can turn into a tuning and latency tax.