Frontier intelligence, without the frontier waste
The model supplies the intelligence. The harness decides how expensive, slow, forgetful, or personal that intelligence becomes. We built this research around one claim: a conversation should be allowed to keep growing without making every answer carry the whole past again.
Use the best model for the answer—not to reread a life it already knows.
The workload
We replayed the same growing conversation through two paths. The conventional path sent the accumulated transcript. Numa used the memory state produced by the running system at the same checkpoints; it was not held to a fixed laboratory-sized prompt. At one million historical tokens, the provider recorded 864,599 input tokens for full replay and 79,122 for Numa: 10.9× less input on that turn.
Provider-recorded input tokens from the production-like growing-memory comparison. Numa's carried record grows too; it simply grows much more slowly than transcript replay.
The wait
We then ran one deliberately simple Opus 5 check: the exact same prompt was sent through both paths at each matched snapshot, and we measured the time until the first visible answer token. At 11K and 31K, Numa was 1.4× faster. The gap grew to 2.0× at 99K, 5.0× at 250K, 4.0× at 497K, and 5.7× at one million — 1.40 seconds rather than 7.96.
Opus 5, one paired observation per path and snapshot. This is an intentionally legible scaling check, not a p95 latency study. Absolute times vary with geography, provider load, network conditions, and cache state; the observed pattern is that full replay slows much more as history grows.
The bill
To make the economic scale concrete, we audited one team member's 3.5-year ChatGPT export: 7,604 prompts across 1,140 chats, 10.4 prompts per active day, and 177 answers in the average active month. Once that relationship had crossed one million visible tokens, preserving it through rolling full-context replay at Fable 5 API prices would model to $1,775 for that active month. Repricing the complete token ledgers from two mature Numa memory runs at the same no-cache price sheet gives $20–$34.
Counterfactual API-equivalent, not a ChatGPT invoice or paid Fable run. Both sides use $10/M input, $50/M output, and no cache. Isolated chats cost less because they reset and do not preserve the same cross-chat relationship.
What survived
Efficiency is only useful if the relationship survives it. On LongMemEval, our strongest observed development run reached 92.4% with DeepSeek Chat in non-thinking mode. On the controlled 500-answer comparison under one blinded shared rubric, Numa scored 79.2% and Mem0 87.2%. On the same 120 BEAM 1M answers under one strict judge, Numa scored 56.7% and Mem0 64.2%. We compete; we do not pretend to win every table.
These controlled comparisons share answer sets and graders within each benchmark. They are separate from vendor headline tables, whose answer models, retrieval budgets, and protocols differ.
The most legible result is also the simplest. In the saved BEAM 1M output, the first message fixed four versions: Python 3.10, FastAPI 0.85, PostgreSQL 14, and React 18.2. 1,779 messages later, Numa recovered all four when asked. Both the candidate judge and the benchmark judge marked the answer correct.
The aim is not fewer tokens. It is a relationship that can keep going.
Memory can be switched off at any time. In that state, the conversation is deliberately isolated: no personalized answer, no inherited assumptions, just an outsider's view. Continuity should be useful, not compulsory.