// CASE STUDY 02
ENGINE // GITHUB.COM/CLUESTATE
OPEN ARCHITECTURE
ROLE: CREATOR & ARCHITECT
ClueState: Evidence-Backed State for Production AI.
“Turning unstructured human conversations into verifiable, schema-driven state without losing source provenance.”
CORE PARADIGM
State & Provenance Layer
Separating language extraction from durable state storage.
KEY INNOVATION
Immutable Evidence Links
Every attribute links back to exact raw conversation snippets.
CONFLICT MODELING
Confidence & Decay
Explicit handling of outdated, competing, or superseded claims.
// 01. THE CRITICAL VULNERABILITY IN LLM MEMORY
Why Text Summaries Fail in Production Software
When developers build AI applications that require long-term user or customer memory, the standard practice is naive: re-prompting the LLM with conversational history, or storing periodic text summaries in a vector database.
This breaks down rapidly in real-world systems. Unstructured summaries lose the original source evidence. When a user changes their mind or updates their preferences, past summaries contradict current facts. Most critically, downstream software components—such as database queries, permission gates, or recommendation engines—cannot reliably query an unformatted prose paragraph.
// 02. THE CLUESTATE ARCHITECTURE
From Unstructured Signals to Auditable Schemas
ClueState introduces a strict separation between extraction and state storage. Rather than trusting LLMs to hold state implicitly, the engine acts as an intermediary pipeline:
- Predefined Schema Enforcement
Conversations, notes, and activity streams are ingested and extracted against strict domain attribute schemas. Candidate facts must match validated types rather than free-form text.
- Immutable Provenance Anchors
Every extracted claim retains a permanent reference to its source evidence: transcript timestamp, message ID, or note offset. An operator or audit process can immediately inspect why the system believes a fact to be true.
- Confidence Modeling & Superseding Facts
Facts are assigned confidence scores based on language ambiguity and source authority. When newer interactions provide conflicting claims, the engine resolves contradictions through explicit versioning and decay rules.
// ARCHITECTURAL ANGLE
Rigorous Data Modeling for Applied AI
ClueState reflects my fundamental belief about applied AI: the real value is not in prompt tricks, but in the systems engineering that anchors probabilistic model outputs into durable, auditable software structures.
APPLICABILITY
Personalization & CRM State
Ideal for retail customer memory, SaaS user preferences, and high-compliance support systems.