ARCHITECTURE

Engineered for cognition

A purpose-built engine designed around memory, state, and context — not adapted from traditional database architectures.

How Durbl is structured

Durbl is organized in layers, each with a specific responsibility. Data flows from the SDK through the engine's processing pipeline, into the storage layer, and back out through the context runtime.

SDK & API Layer

Your application communicates with Durbl through our typed SDKs (Python, TypeScript) or raw REST API calls.

Processing Engine

The cognitive runtime handling data flow, intelligence, and lifecycle events natively without external dependencies.

Write Pipeline

Classifies, scores, and connects incoming cognitive data

Context Runtime

Assembles goal-aware context frames from raw memory

Lifecycle Engine

Reinforces, decays, drops, and manages memory health

Intelligence Kernel

Discovers hidden patterns natively, detects anomalies automatically, and strictly monitors overall systemic memory health.

Storage Layer

Purpose-built, temporal storage with semantic search indexing embedded, relationship tracking, and lifecycle journaling.

Write Pipeline

Every piece of data flows through a multi-step processing pipeline before it becomes memory.

01

Validate

Incoming data is cleaned, validated, and structured into a consistent format.

02

Classify

The engine determines what type of memory this is: episodic, semantic, preference, procedural, or task-related.

03

Score

Each memory receives scores for importance, confidence, and durability — determining how long it persists and how easily it's recalled.

04

Represent

Content is converted into a semantic representation for intelligent retrieval.

05

Route

Based on scores and type, the memory is directed to the appropriate storage tier.

06

Connect

The engine discovers and creates relationships to related memories, entities, and states.

07

Signal

If the new data triggers any patterns, anomalies, or state changes, signals are emitted.

Context Runtime

Context is computed, not fetched.

When your application requests context for a specific goal, the runtime executes a multi-step assembly process to deliver the most relevant, diverse, and policy-compliant context.

Gather

Candidate memories are gathered using semantic similarity, entity scope, and temporal relevance.

Rank

Each candidate is scored by relevance to the stated goal, combined with its importance and recency.

Diversify

The engine ensures variety across memory types — avoiding redundant context.

Assemble

The final context frame is structured and optimized for consumption by language models or application logic.

Lifecycle Engine

Memory that maintains itself.

Without active lifecycle management, AI memory becomes noisy, outdated, and expensive. Durbl continuously maintains memory health.

Reinforce

Accessed or confirmed memories become stronger and more durable.

Decay

Unused memories gradually lose priority over time.

Merge

Similar memories are intelligently combined.

Summarize

Large volumes of episodic memory are condensed while preserving key information.

Archive

Valuable but inactive memories move to efficient long-term storage.

Forget

Memories below minimum thresholds are permanently and intentionally removed.

Intelligence Kernel

Discovering meaning across time.

The engine continuously monitors all data flows, identifying significant events without prompts. It surfaces 4 core signal types to your application.

Patterns

Repeated behaviors, consistent preferences, or cyclical events discovered over time.

Anomalies

Events that significantly deviate from an entity's established baseline or past behavior.

Milestones

Major state thresholds, goal achievements, or lifecycle events.

Conflicts

Contradictions between new cognitive events and firmly held historical memory.

The 6 Primitives

Durbl's data model is built on 6 cognitive primitives.

MemoryObject

The atomic unit of memory. Carries content, semantic embeddings, quality scores, and lifecycle metadata.

StateObject

Versioned snapshots of entity state. Immutable history with temporal queries.

ContextFrame

Assembled context for a specific goal. The bridge between memory and reasoning.

MemoryLink

Typed relationships between memories. Enables knowledge graph traversal.

CognitiveEvent

Raw input events that flow through the write pipeline.

PolicyObject

Governance rules for retention, decay, access control, and compliance.

Design Principles

Engine-first

Durbl owns its data model, processing pipeline, and storage layout. It is not a wrapper over existing databases.

Memory as a first-class citizen

Memory is not an afterthought or a feature. It is the core primitive.

Time is native

Temporal ordering, change tracking, and historical queries are built into the foundation.

Lifecycle matters

Data without lifecycle management becomes noise. Durbl maintains memory health automatically.

Intelligence emerges

The engine observes, discovers, and signals — it doesn't just store.

Performance

Durbl is engineered for low-latency, high-throughput AI workloads.

OperationTarget Latency
Memory write< 1ms
Semantic recall< 5ms
Context assembly< 15ms
State read< 0.1ms

* These are engineering targets for V1. Actual performance depends on deployment configuration and workload.