AI Forgets Everything. We Measured What That Costs. Everyone is arguing about how much energy AI uses. Almost nobody is asking how much of that energy is spent repeating itself. Here is the pattern. An AI coding session builds up context: it reads your files, searches your codebase, learns your project. Then the session ends and all of that is thrown away. Tomorrow it reads the same files, runs the same searches, asks you the same questions. The model didn't get smarter overnight. It got amnesia. We measured what that forgetting costs, on real data, and we're publishing the receipts: the full 79-row measurement table and the conversion script ship alongside this piece. The Dataset One power user, one machine. The core measurement: 14 full session transcripts from the last six weeks, totaling 1.27 GiB, containing 24,506 API calls and 12.6 billion input-side tokens, all pulled from API-reported usage fields, not vibes. Behind that: 5,289 working sessions and roughly 262,000 messages persisted, encrypted, to the BSV blockchain since February. To be precise about what the chain proves: the sessions exist, timestamped and immutable. The token counts themselves come from API-reported usage fields in the local transcripts, not from the chain. One detail from that dataset worth sitting with: 97.7 percent of all input-side tokens were cached re-reads of context the model had already seen. Modern AI work is mostly the model looking at the same material again. Cached re-reads are far cheaper than fresh computation, which is why our energy math excludes them entirely. The expensive part is context that has to be rebuilt from scratch, and that is the part we measured. The Experiment the Harness Ran for Us These transcripts contain 79 measured context-replacement events. They are Claude Code's own compaction boundaries: the harness fills its context window, distills the session into a summary, and swaps the huge accumulated context for a compact one. Each event has exact before and after token counts from the API. Compaction is the in-harness, single-session version of the same move; Indelible's on-chain restore is the permanent, cross-session version. These boundaries are the ground truth of what context replacement is worth, whoever performs it. The median event replaces 971,805 tokens of live working context with 71,266 tokens of restored context. That is a measured 12.3x compression: 92.7 percent fewer tokens recomputed. The harness's distilled summary doing the carrying is about 4,600 tokens, roughly 208x smaller than the context it stands in for. Indelible's on-chain restore injections are smaller still: 0.35k to 2.6k tokens per restore call. Compare that to what a cold rebuild actually reads: 7.9 million tokens of file reads and search results accumulated in one session's re-derivation (6.7 million in a second session, so it replicates). A restore is three to four orders of magnitude less reading than doing it over from scratch. Tokens to Watts Short-prompt chat medians (Google's measured 0.24 Wh, OpenAI's reported 0.34 Wh) don't apply to million-token contexts, so we used Epoch AI's published long-context points: 2.5 Wh at 10k input tokens, 40 Wh at 100k. That gives a band of 0.25 to 0.40 mWh per input token that actually has to be recomputed. This is an estimate, clearly labeled, and we applied it only to tokens forced through fresh computation. We did not energy-price the 12.3 billion cached reads, because no verified published figure quantifies the cache discount, and we'd rather undercount than invent. Run the arithmetic. Rebuilding a median boundary cold: 225 to 360 Wh. Restoring instead: 18 to 43 Wh. Net saving per return-to-work event: roughly 183 to 342 Wh. That is the energy of 540 to 1,400 median chatbot queries, per event, at the published per-query figures. Said plainly: restoring costs about a tenth of rebuilding. Roughly ninety percent of the energy of coming back to work, avoided (81 to 95 percent across the band). Of that moment, to be clear, not of all AI energy. That precision is the point. This user hits 1.1 such boundaries per day. Over ten weeks the measured window shows 72 million tokens of replaced context against 7.1 million tokens of restore cost, about 16 to 26 kWh net avoided. Annualized: roughly 73 to 137 kWh of inference energy avoided per year, for one heavy user, net of the restore calls themselves. The Caveats, Stated Plainly Because that's the whole point. This is one power user, not a population. Two mega-sessions dominate the data. Do not multiply these numbers by "all AI users." The Wh conversion is an estimate built on Epoch's modeled figures, extrapolated past their stated 100k range. If attention cost is superlinear, we're undercounting the rebuild side. If provider caching improves, both sides get cheaper. These savings cover inference only. Inference is estimated at 80 to 90 percent of AI compute (an industry estimate, no single primary study). Training, hardware, and network are excluded. And no, we are not using the "30 percent of datacenter energy" number going around this thread. We couldn't verify it, so it's not in our math. We also excluded an IEA figure we couldn't reach at the primary source. If we can't check it, we don't claim it. The token counts, though, are not estimates. Every before-and-after context size is exact API-reported usage, and the session history backing them is written to a public blockchain where anyone can verify it exists. The Point The industry is debating how to power AI's future. Fair. But a system that re-derives a million tokens of context it held yesterday is not a compute problem, it's a memory problem, and memory is cheaper than recomputation by a factor we can now put a number on. Indelible stores AI working sessions encrypted on-chain so a returning session restores its memory instead of rebuilding it. That's it. That's the product.