A Markdown wiki outscored every AI agent memory product we benchmarked

I'll pull the source article so the paragraphs stay grounded in real claims and numbers rather than inventing detail.Verging Labs published Agentic Memory Index v0.1 (August 2026), an independent head-to-head of tools that store and retrieve facts for AI agents. On that index, Karpathy Wiki took first place with a score of 98.5 and about 98 percent correct answers across task types, ahead of every dedicated agent-memory product in the field. The pack behind it included Mitosis Cortex (96.9), gbrain (92.9), Hyperspell (92.4), Mem0 (92.3), Anthropic Memory (84.0), Supermemory (77.1), Zep (75.1), and Claude Code built-in memory (67.7). The result is not a marketing claim from a vendor: Verging Labs states that no provider pays for placement or scores.

The benchmark scores tools the same way on accuracy, cost per successful answer, end-to-end answer latency, and failure modes, with the same agent model on every memory backend. Task-type breakdowns cover direct recall, updated facts, thread growth, synthesis, long-term retention, and a false-memory check; long-term retention is probed at five checkpoints across 56 working sessions. Karpathy Wiki also posted the lowest incorrect-answer rate in the failure table at 1.0 percent, versus mid-single digits for several commercial options and 23.5 percent and 33.0 percent for Anthropic Memory and Supermemory. On retrieval, it averaged 2.7 seconds per answer and only 1.6 lookups per session—fewer round trips than peers that needed five to eight lookups.

For engineers shipping agents, the headline is that a plain Markdown wiki, treated as an open method with zero list price, beat purpose-built memory APIs on overall correctness and failure rate. That matters when you are choosing between a managed memory service and a file-backed knowledge store the agent can read and write with ordinary tools. Built-in product memory is not a free win either: Claude Code’s bundled memory sat last on the index at 67.7 with a 39.5 percent incorrect-answer rate, so “whatever ships in the IDE” is a measurable design choice, not a default to trust blindly.

Market context is crowded and expensive at the top of the paid tier. List pricing (as of 27 July 2026) ranges from free or open-core options (Karpathy Wiki, gbrain, Anthropic Memory, Supermemory core) through Mitosis Cortex at $7.99 per month and Mem0/Supermemory around $19 per month up to Zep at $125 per month. Cost per 1,000 successful answers still looks different from sticker price: Mem0 was cheapest in that metric at $341.42, while Karpathy Wiki was the most expensive at $568.93—driven mainly by agent token spend rather than a subscription. Speed and onboarding also diverge: Zep needed 9.1 minutes to onboard and 162.7 seconds to become ready; Karpathy Wiki onboarded in 1.4 minutes with local readiness.

Watch two numbers when you re-run or re-score your stack: correct-answer rate by task type you actually care about (synthesis and long-term retention often diverge from pure recall), and fully loaded cost per 1,000 successful answers, not monthly list price alone. If your agent already works well with local files, a Markdown wiki is a serious baseline to beat before you pay for a specialized memory product; if you need multi-agent scale, multi-tenant APIs, or managed graph-style memory, compare Mitosis, Mem0, and peers on the same failure modes Verging Labs publishes, not on category claims.