Introduction

Structured, customizable, and managed memory for agency economics.

How It Works

Want to discover (almost) all our secrets? Read here.

Benchmarks

Validated on the LoCoMo dataset.

Quickstart

Start with MemoryModel in guided steps.

Examples

If you want examples of usage.

Memory Model is a fully managed Adaptive Intelligence Platform designed to solve context retention for LLM applications. Unlike static vector databases that rely solely on similarity search, Memory Model operates as an active orchestration middleware. It combines a Schema-Agnostic Storage Engine with an Adaptive Retrieval System that autonomously manages data ingestion strategies, query routing, and parameter self-optimization.

Data Strategy & Ingestion

User-Defined Schemas

The platform utilizes a Schema-Agnostic approach. Instead of storing generic text chunks, it operates on Specialized Memory Nodes. Users define custom structures via the Management Console to capture specific attributes alongside unstructured data.

Event-Driven Ingestion Pipeline

The architecture employs an asynchronous processing model. Upon ingestion, memory nodes enter a processing queue to undergo a deterministic two-stage transformation before final storage:

1. Multi-Stage Semantic Enrichment: The system applies bidirectional semantic expansion. It injects implicit context (themes, related concepts) into the node to maximize vector overlap, eliminating terminology mismatches during retrieval.
2. Shift-Left Temporal Resolution: Relative time references (e.g., “next Friday”) are resolved into absolute ISO 8601 timestamps. This creates deterministic indices, converting complex temporal reasoning into precise date-range lookups.

Virtual Knowledge Graph

Beyond storing isolated vectors, the architecture maintains a logical graph structure. By analyzing shared entities and temporal proximity, the system links disparate memory nodes into a cohesive network. This topology allows the system to traverse relationships (e.g., connecting “Health” nodes to “Shopping” nodes) and generate Synthesized Insights—higher-order nodes representing behavioral patterns that would be invisible to standard similarity search.

Retrieval Logic

The core differentiator is the Adaptive Retrieval System. The platform abandons “one-size-fits-all” searching in favor of Intent-Based Routing, classifying every query into one of four execution strategies:

1. Direct Match Strategy: Executes precise key-value lookups for specific identifiers (IDs, filenames), bypassing vector search for O(1) performance.
2. Entity Anchor Strategy: Activated when the intent focuses on specific Named Entities, utilizing the Knowledge Graph to fetch content tied to a subject regardless of semantic phrasing.
3. Temporal Range Strategy: Converts time-bound queries into deterministic date-range filters (leveraging the pre-resolved ISO 8601 timestamps).
4. Adaptive Vector Search (Semantic): Handles abstract or conceptual queries. Uniquely, this strategy utilizes Centroid Analysis (measuring distance from the user’s semantic center of mass) to dynamically select an operating mode:
   • META Mode: Broad, exploratory search (High-k).
   • SPECIFIC Mode: Narrow, precision-focused search (Low-k).

Outputs from these strategies are aggregated via a Fusion Layer responsible for deduplication and final ranking.

Continuous Optimization

The platform replaces manual configuration with Closed-Loop Optimization. Background processes, governed by Control Theory principles, continuously analyze retrieval telemetry (Precision/Recall). The system automatically adjusts similarity thresholds and ranking weights to adapt to evolving user patterns without human intervention.