Savings calculator
What memory saves, with the working shown.
Model spend per call is a sum of token blocks times a price. A memory layer replaces two of those blocks and leaves the rest alone. Put in your workload; every number below carries the line it was derived from.
Annual saving on model spend
$0per year
Above 80%. This configuration lands where the industry's headline numbers live, and a technical buyer will challenge it. The defensible published range is 55 to 75% against full-context replay and 40 to 55% against tuned RAG with caching.
Input tokens per call
0%
Input cost per call
0%
Total bill
0%
Input tokens per call, by block
Only the two middle blocks change. Hatched blocks are served from the prompt cache at the hit rate you set.
Not addressable: system, tools, query
History, becomes the compiled profile
Retrieved context, becomes facts + excerpts
Today
With memory
Where the money goes
Per call and per year. Upkeep is memory's own model cost and is charged, not hidden.
| Line | Today, call | Memory, call | Today, year | Memory, year |
|---|
The larger prize: one model class down
Assumptions and sources
Prices
First-party Claude API list prices, USD per million tokens, verified . Change the model pickers to compare tiers; edit the table in the page source if your contract differs.
| Model | Input | Output |
|---|
Prompt caching
- Cache reads bill at 0.1x the input price, cache writes at 1.25x (5-minute TTL). A block that is cacheable bills at hit × 0.1 + (1 − hit) × 1.25. A hit rate of 0 means caching is switched off and bills 1.0x.
- Cacheable in both scenarios: the system prompt and tool definitions.
- Cacheable today only under full replay: the conversation history, because it is a contiguous prefix. A rolling window or retrieved chunks shift every call and miss.
- Cacheable with memory: the compiled profile block, which only changes bytes when a fact changes.
- Never cacheable: retrieved chunks, facts and excerpts, the query. This is why caching and memory stack rather than compete.
Memory upkeep
- Remembering, indexing and recall use no model call. The one call is an asynchronous distillation per session, run on the model you pick for the memory scenario, and amortised over the calls in that session.
- Kernel storage and compute are ordinary cloud costs and are left out on both sides.
Deliberately not in the calculator
- Accuracy and stale-answer rate. The published trade-off is that memory systems give up a few points of accuracy for cost. Breaking that trade-off is the claim we are measuring; it goes in when the benchmark result exists, not before.
- Latency. Published memory systems cut p95 from about 17 s to about 1.4 s against full context. It changes concurrency and user experience, but turning it into dollars needs your infrastructure bill, not ours.
Where the ranges come from
- Mem0, "Building Production-Ready AI Agents with Scalable Long-Term Memory" (arXiv 2504.19413): full context at about 26,000 tokens per query and 73% accuracy; their memory at about 1,800 tokens and 67%. The often-quoted 90% is the footprint compression in that comparison, not a task-level saving.
- MemSyco-Bench: average input tokens across systems, with full dialog and naive RAG near 1,100 and Mem0 near 430, about 62%. Memory systems mainly reduce input length; output length barely moves.
- A controlled coding-agent benchmark reported 15 to 28% end-to-end savings, the honest floor when the baseline is already tuned and cached.