AI middleware that understands emotion.

Most memory layers remember facts. Belcore also carries the feeling between them — a deterministic emotional state that advances every turn, so replies answer the relationship instead of the last message. Long-term recall, token cost, and security are how we make that hold up in production.

This engine runs SoulLink, our own companion app, shipped on iOS and Android. It was not built for a benchmark and then pointed at users: it came out of live traffic, and the numbers below were measured afterward.

The claim

Emotion engine — what we lead with

Longing, sulkiness, warmth: a state that advances every turn and enters the prompt, so a character who was hurt yesterday still sounds hurt today. Deterministic — same inputs, same state — so you can inspect why a reply came out the way it did.

No shared industry benchmark exists for emotional continuity, so we publish no comparative number here.

96.79%

Long-term recall — the evidence

Emotional continuity is worthless if the facts drop. Across 500 multi-session conversations, the context Belcore assembles contains the answer-bearing turn.

Retrieval recall (gold coverage) on LongMemEval_S: 96.79% overall, 95.95% multi-session. This measures whether the assembled context contains the answer-bearing turn — a retrieval metric, not end-to-end task accuracy.

93.88% ↓

Token cost — the evidence

Carrying a relationship cannot mean re-sending the whole transcript. Measured end to end on a public long-conversation benchmark, not projected from a demo.

6,671 input tokens per call vs 109,079 uncut. LongMemEval_S, 500 questions.

3 modules

Security — the evidence

A persona is only consistent if it cannot be talked out of itself. Injection defense is deterministic and server-side, never delegated to the model, and a detected attack is deflected in character rather than surfaced as a failure.

Provisional patent applications filed. Not granted patents.

Any model · One API

OpenAI, Gemini, or your own stack. Belcore is a layer in front — no migration, no model lock-in.

ALIGNED PRICING

No verified savings, no fee.

Savings are verified per call: baseline (the full conversation sent uncut) minus actual (what Belcore sent). We bill 10% of that difference — nothing else.

10%

Belcore receives 10%

90%

You keep 90%

1.Benchmark: LongMemEval_S, 500 questions, ~109k tokens of prior conversation per question. Fixed seed, frozen harness; token counts measured with the o200k_base encoding.
2.Latest 500-question run: overall retrieval recall (gold coverage) 96.79%, multi-session recall 95.95%. Gold coverage asks whether the retrieved context contains the answer-bearing turn; it is a retrieval metric, not end-to-end accuracy. Figures from other memory vendors may use different metrics and benchmarks.
3.93.88% is the input-token reduction against sending the full conversation uncut. Mem0 publicly claims ~90% token savings, but on a different dataset and methodology — the two figures are not directly comparable.
4.Security modules are at provisional-application stage. No patent is granted, and nothing here should be read as an issued patent. The emotion engine is not benchmarked against competitors because no shared benchmark exists.