Failure 01
Topic domination
High-frequency topics crowd out everything else. The grocery-list mention from yesterday and the family-death disclosure from last month get the same retrieval treatment.
Mem0, Zep, MemoryBank, and the published literature all assume more engagement equals stronger memory. KAPEX inverts the curve. More processing → faster decay. Unresolved content persists. Resolved content recedes. That's the patent.
The problem with current memory systems
User mentions their father 50 times → father memories dominate every context window. Indefinitely. This models obsession, not understanding. Four failure modes follow.
High-frequency topics crowd out everything else. The grocery-list mention from yesterday and the family-death disclosure from last month get the same retrieval treatment.
Resolved grief and active crisis look identical in retrieval. Both have high recall counts. Both surface. The system can't tell them apart.
"I had Thai food for lunch" and "my father just died" get treated as equal-weight memory nodes. No scoring at write-time means no scoring at retrieval-time.
Without write-time scoring, retrieval fills the budget with stale, undifferentiated nodes. The model gets verbosity instead of signal.
What KAPEX does differently
Send a conversation turn to KAPEX, get a scored memory context block back, inject it into your LLM prompt. The core difference: processing-modulated decay. More processing means faster decay. This isn't a parameter tweak — it's a fundamentally different model.
Your app passes the user input and your LLM's response to KAPEX. KAPEX scores, stores, and updates the memory graph — without changing your model.
At query time, KAPEX returns a token-budgeted context block of the highest-salience memories for that user. Three retrieval channels compete for the token budget.
Paste the context into your system prompt. The model now remembers — across sessions, across products, across providers. Works with Claude, GPT, Gemini, Llama, or your own.
Architecture at a glance
The architecture is structured, not a single black-box embedding model. Each engine is independently testable, swappable, and patent-pending.
12 signal dimensions computed at write-time. Detail →
Detects when what someone says doesn't match how they say it — the patent-worthy signal under NDA.
Domain → entity → facet. Not flat vectors. Detail →
Memories decay toward category-specific minimums. They never fully disappear.
Salience + recency + constraint, token-budgeted. Detail →
Crisis sentinel, anti-fabrication, PII, prompt-injection. Detail →
How to evaluate
The fastest way to evaluate KAPEX is a controlled side-by-side against your current memory system. Here's the protocol we use with pilot prospects.
Run the same 20-message conversation through KAPEX and through your current memory system. Watch what happens at message 15 when the user has clearly moved past a topic discussed heavily in messages 2–6. In other systems, messages 2–6 still dominate context. In KAPEX, they've decayed proportionally to how much the user processed them.
If you want a pre-built eval harness with our internal scenarios, ask during pilot onboarding — we'll send it under mutual NDA.
Moat metrics
From our internal benchmark suite. Methodology, exact scores, and the full result set are shared under mutual NDA.
Experience KAPEX live
Sign up for the free KAPEX beta and see salience-scored memory in action — no NDA, no commitment.
30-day pilot, full feature set, 100 users. We'll send the eval harness under NDA.