A neutral look at token efficiency and governance

• Both are agent-memory systems that score at the top of LongMemEval (Hindsight reports 91.4%; Zep reports 90.2%) — a single benchmark won't decide your choice.
• Hindsight's 91.4% is measured at an 8,192-token retrieval budget — about 2× the ~4,408 tokens Zep uses for 90.2% on the same benchmark. That extra context is fed to your answer LLM on every call, so at the headline numbers Hindsight costs roughly double the memory tokens per query — real money and added latency at scale.
• The other production differences are enterprise governance, managed operations, proven scale, and deployment control, where Zep is purpose-built. (Both systems do temporal reasoning and provenance — see the table.)
• Hindsight is MIT open source with a biomimetic memory model that you self-host and operate; its managed/hosted offering is newer. Zep ships a managed Context Lake with SOC 2 Type II, HIPAA, and BYOK/BYOC today.
• Pick by need: open-source self-hosting and a human-memory-style model → Hindsight; governed agent memory at enterprise scale, at lower per-query context cost → Zep.

What Vectorize Hindsight is. Hindsight (vectorize.io, GitHub) is an open-source (MIT) agent-memory system from Vectorize. It organizes memory with biomimetic structures — World (facts), Experiences (the agent's own history), and Mental Models (learned understanding formed by reflecting on raw memories) — integrates in about two lines of code, and reports 91.4% on LongMemEval. It offers self-hosting (Docker/embedded), and Vectorize is building a hosted cloud version for managed, production features.

What Zep is. Zep is a dedicated, managed agent-memory platform — the Context Lake. It builds bi-temporal context graphs (via the open-source Graphiti) in which every fact carries a validity window and provenance, so the agent reasons over what's true now vs then and can audit any answer to its source. It serves millions of graphs at sub-200ms p95, governs memory in the data layer (ABAC, retention, audit), and deploys managed, BYOK, or BYOC. Zep reports 90.2% on LongMemEval and 94.7% on LoCoMo (results); architecture in the Zep paper.

The published comparison isn't a controlled, matched-backbone head-to-head — by Vectorize's own account. Their repo states that only Hindsight's score was independently reproduced (Virginia Tech, The Washington Post) and that “other scores are self-reported by software vendors.” The Zep figure they cite (71.2%) is Zep's 2025-paper number, not Zep's current 90.2% — so the comparison pairs Hindsight's reproduced Gemini-3 Pro result against Zep's older self-reported figure.

Mechanically the two systems are similar: Hindsight extracts facts with an LLM on ingest (“retain”), and on recall runs vector + BM25 + graph + temporal retrieval merged with reciprocal-rank fusion and a cross-encoder reranker, trimmed to a token limit — the same shape as Graphiti/Zep. (Its “LLM-free recall” claim applies to recall, not ingest.) Hindsight publishes a separate speed/cost benchmark but no per-query latency/token figure in the accuracy table.

Context size (from Hindsight's own benchmark code). Their LongMemEval runner defaults to an 8,192-token retrieval budget at Budget.HIGH (thinking_budget=500), with the answer model at high reasoning effort. That's roughly 1.9× the ~4,408-token context Zep reports on the same benchmark — so the 91.4% is achieved with about double the retrieved context (and a top-tier backbone). On a token-matched basis the gap narrows or reverses. (Hindsight's leaner LoCoMo quality benchmark uses a 4,096-token “low” budget.) Read accuracy alongside latency and tokens, on a matched backbone, against Zep's current numbers. (Hindsight paper: arXiv 2512.12818.)