Buyer’s guide · 17 September 2026
There is no honest one-number leaderboard. LoCoMo scores now depend heavily on reader model, judge, reranking and retrieval budget. Choose the memory architecture that matches your deployment constraint first, then benchmark it on your workload.
Mycelic is the fit when raw memory must remain local and you care about reconstructing claims across independent holders with explicit lineage. The current research harness measures an ~65 ms mean retrieval stage for the shipped local_pairs configuration on its 282-question single-hop slice, but this is not a vendor-comparable p50 or p95. It is still a research engine, not a managed memory service.
Zep currently publishes one of the clearest production-oriented LoCoMo reports: 94.7% on 1,540 questions, with 87 ms p50 and 155 ms p95 retrieval, plus a 5,760-token median context. Its current Flex plan starts at $125/month. Those numbers are Zep’s own 2026 methodology, so compare them within that protocol rather than against an older third-party score.
Mem0 has a free Hobby tier, a $19/month Starter tier and a $249/month Pro tier that includes graph memory. Its updated 2026 algorithm reports 92.5% LoCoMo with 0.88 s p50 retrieval. It is a practical default to evaluate when you want a mature general-purpose memory layer and do not need Mycelic’s local-first lineage model.
Hindsight publishes both an open benchmark harness and an open-source memory system. Its latest AMB post reports 92.0% LoCoMo in single-query mode; its docs describe typical recall at 50–100 ms without reranking and 200–500 ms with reranking. Cloud pricing is granular: retain, recall, reflect and storage are metered separately.
AgentCore Memory is the operational choice for teams that value AWS integration over a standalone memory startup. Pricing is consumption-based: $0.25 per 1,000 short-term events, built-in long-term storage at $0.75 per 1,000 records per month, and $0.50 per 1,000 long-term retrievals. AWS does not publish a directly comparable LoCoMo result in the material checked for this guide, so we do not manufacture one.
Local/private research: evaluate Mycelic. Low-latency managed cloud: evaluate Zep. General developer platform and cheap entry: evaluate Mem0. Open memory research and metered cloud: evaluate Hindsight. AWS-native enterprise: evaluate AgentCore Memory.
Then run the finalists on the same questions, hardware region, reader model, judge, top-k, reranker policy and token budget. Until that test exists, “94% vs 92%” is usually a protocol comparison disguised as a product comparison.
For a controlled LoCoMo view, use the Hindsight comparison table where systems were run under one published harness: Backboard 90.00%, Hindsight Gemini-3 89.61%, Hindsight OSS-120B 85.67%, Hindsight OSS-20B 83.18%, Memobase 75.78%, Zep 75.14%, Mem0 Graph 68.44%, Mem0 66.88%. Mycelic (73.8%, own leakage-free harness, not the Hindsight protocol) is shown alongside it for reference, not as a same-protocol result. This is not the same protocol as each vendor’s latest 2026 headline number.