How MemQ Works
MemQ is not a single database. It is a multi-tier memory fabric that combines fast inference, background reflection, and tiered storage to mimic human recall capabilities without revealing sensitive data.
System Architecture Overview
The MemQ system is broken down into two distinct halves: the AI Processing Layers that handle raw ingestion, and the Memory Tiers that persistently store data.
On smaller screens, scroll horizontally to inspect the full routing map.
AI Processing Layers
1. Inference & Embedding
Converts raw text, code, or logs into high-dimensional vectors and structured JSON entities in real-time.
2. Distillation & Reflection
Asynchronous background processes that generalize concepts, resolve contradictions, and repair memory structures.
3. Reasoning & Orchestration
The active execution engine that plans retrieval strategies and bridges contexts for multi-step agent tasks.
Memory Storage Tiers
HOT Tier
Episodic session-level context. Highly volatile, ultra-fast cache.
WARM Tier
Semantic project-level similarity search via high-dimensional vector indices.
COLD Tier
Relational, permanent structured knowledge graph representing entities and relationships.
JOURNAL Tier
Immutable sequential reasoning and audit logs used for compliance and replay.
Decay & Reputation Mechanism
Unlike static databases that grow infinitely, MemQ implements a self-regulating Decay & Reputation cycle. This ensures agents do not drown in irrelevant or outdated context over long lifespans.
How Memory Degradation Works
Reinforcement: When a memory is recalled from the WARM or COLD tier and successfully used to solve a problem, its reputation_score increases, making it more likely to be retrieved in the future.
Decay: Memories that are rarely recalled, or are explicitly marked as leading to failure paths, experience score decay. Once a threshold is crossed, they are compressed and archived.
Pruning: Severely degraded memories are pruned from the HOT and WARM tiers, leaving only a lightweight trace in the COLD graph or JOURNAL audit ledger.