Patent pending·Self-hosted

What's inside the container.

KAPEX ships as a single Docker container. Inside: a memory engine built for the realities of production AI — salience scoring, processing-aware decay, multi-channel retrieval, a safety pipeline, and the governance to run it in regulated environments. The page below is the public overview. The technical reference lives behind NDA.

Under NDA Signal weights, formulas, module architecture, and benchmark methodology are shared only with pilot teams under mutual NDA. Request NDA
01Scoring engine

Score what matters at write time.

Every memory node receives a composite legitimacy score the moment it is ingested — not at query time, not after the fact. The scoring runs across multiple linguistic dimensions designed to separate enduring relevance from passing chatter.

What it does
  • Multi-dimensional analysis. Each memory is evaluated against multiple independent linguistic signals that capture significance from different angles.
  • Cold-start ready. Meaningful scores from message one. No retrieval history or warm-up period required.
  • Computed at ingestion. Scoring happens at write time, so retrieval is fast and ordering is deterministic across sessions.
Why it matters
  • Vector stores treat all memories as equal-weight nodes. KAPEX treats them as scored signals and that's what makes retrieval coherent over time.
  • Significance ≠ frequency. Frequent doesn't mean important. KAPEX measures the difference.
Under NDA

Specific signals, signal counts, weight derivations, normalization scheme, and the legitimacy gap mechanism are shared under mutual NDA during pilot onboarding. The scoring engine is the protected core of KAPEX's IP.

02Decay engine

Processing-modulated decay. Patent pending.

The crown-jewel innovation. When a user actively processes, reflects on, or resolves a topic, that topic's prominence decreases over time. Unresolved topics persist. This is the mathematical inverse of every published memory system — and the reason context windows feel coherent over weeks instead of cluttered with stale repetition.

Behavioral outcomes
  • A user working through grief over multiple sessions: grief context gradually fades from active injection, making room for current concerns.
  • A user mentioning their dog casually across many sessions: the dog stays at moderate salience — never deeply processed, just consistently present.
  • A user who disclosed a family conflict once and never returned to it: the conflict persists at high salience. Unresolved.
What's adjustable
  • Category-specific decay. Different memory types fade at different rates. Identity persists longer than logistics. Configurable per tenant on Growth and above.
  • Residual floors. Memories never fully disappear; each decays toward a category-specific minimum. Foundational disclosures retain permanent baseline salience.
  • Bifurcated decay available on Scale and Enterprise — an advanced patent-pending mode that handles episodic and semantic content on separate trajectories.
Under NDA

The decay model derivation, parameter ranges, and the specific mechanism behind processing modulation are shared only under mutual NDA. This is the patent-protected core of KAPEX.

03Retrieval engine

Token-budgeted context, every time.

Retrieval assembles a memory context block within a token budget. The mechanism balances salience, recency, and constraint-pinned content so no single topic monopolizes the model's working memory.

What you get back
  • Ranked salience. Highest-significance memories get first claim on the token budget.
  • Confidence-gated framing. High-confidence memories are asserted directly; low-confidence ones surface as conversational hooks — the model never states an uncertain memory as fact.
  • Contradiction handling. When retrieved memories contradict, the contradiction is flagged so the LLM can address it gracefully instead of confidently misremembering.
Why this matters

Most "memory" implementations dump a top-K vector match into the prompt and hope the LLM sorts it. KAPEX returns ordered, budgeted, qualified context — the model spends compute on responding, not on triaging stale recall.

Under NDA

Channel definitions, budget allocation logic, expansion strategy, and the multichannel-fusion mechanism (Scale+) are shared under mutual NDA.

04Safety layer

Production safety, identical across every LLM.

An independent safety pipeline that cannot be overridden by memory state, user input, or operator configuration. It runs identically regardless of which LLM you call downstream — the same safety properties apply whether you're on Claude, GPT, Gemini, or your own fine-tune.

What it protects against
  • Crisis indicators. Multi-layer detection routes flagged conversations to safe-handling protocols, including crisis-resource surfacing where appropriate.
  • Fabrication. Multi-layer guards cross-check LLM output against the actual memory graph — the model can't invent memories the user never disclosed.
  • PII leakage. Automatic detection and scrubbing of sensitive identifiers at ingestion.
  • Trigger material. Configurable trigger-content management with graduated sensitivity rather than blanket erasure.
  • Prompt injection. Adversarial-input detection for attempts to manipulate the memory system through prompt engineering.
Governance properties
  • Immutable safety pins. Platform-level instructions that operators cannot weaken at the tenant level.
  • Append-only audit logging. Every safety event is recorded with timestamp, actor, and operation.
  • Jurisdictional awareness. Configurable rules for regulated geographies (e.g., crisis routing per local guidance).
Under NDA

Specific module counts, sentinel architectures, fabrication-guard layer order, and the policy-enforcement mechanism are shared under mutual NDA.

05Entity intelligence

Structured memory around things that matter.

KAPEX builds memory around real-world entities — people, places, projects, decisions — rather than flat vector embeddings. Different facets of the same entity score and decay independently.

Capabilities
  • Cross-session coreference. "My dad," "Father," and pronoun references resolve to the same entity over time, not as separate memories.
  • Independent facets. One entity can hold many independently-scored facets. Sarah-the-coworker and Sarah-the-friend can decay on completely different trajectories.
  • Relational stance tracking. Quality, communication patterns, and sentiment trajectory toward each entity evolve over time.
Why this matters

Flat vector stores tag "Sarah told me about her divorce" and "Sarah and I went to dinner" with the same name and treat them as equal-weight memories. They're not. One is identity-level, one is logistical — and KAPEX scores, decays, and surfaces them independently.

Under NDA

The full entity hierarchy, facet-resolution rules, and the relational-profile tracking schema are shared under mutual NDA.

06Compliance & governance

Built for regulated industries from day one.

Deletion & compliance
  • Per-node deletion. Remove specific memories without destroying the graph — one API call. GDPR Article 17, CCPA, and HIPAA-compatible.
  • Full user deletion. Cascade across nodes, entities, edges, and history.
  • Append-only audit log. Every operation timestamped, actor-tagged, cryptographically chained.
Tenant + deployment isolation
  • Self-hosted by default. The KAPEX container runs in your VPC. Your data never leaves your environment.
  • Multi-tenant isolation (within your environment). Per-tenant graphs, scoring parameters, and decay coefficients. Strict storage / compute / retrieval isolation.
  • Tiered governance. Platform-level safety floors that operators cannot weaken — minimum decay rates, mandatory safety modules, jurisdictional crisis routing.
  • HIPAA BAA & SOC 2 roadmap on Scale and Enterprise. See Security.
The model

Licensed software. Self-hosted by default.

KAPEX ships as a Docker container with an annual license key. You run it in your AWS, GCP, or Azure account. Memory data lives in your database. Sandstone provides the container, the license, and the support — and never sees a single byte of user data.

See pricing Deployment model

Experience KAPEX live

Sign up for the free KAPEX beta and see salience-scored memory in action — no NDA, no commitment.

Try the free beta
Patent pending

See the engines run on your data.

Start the 30-day trial. Full feature set, founder support. Architecture deep-dive under mutual NDA.