LearningX Engine v2.0

The Persistent Mind of DecisionX Agents

Beyond simple RAG. Give your agents a five-layer cognitive architecture to remember facts, learn from context, and evolve with every interaction.

Memory-R1 Pipeline
On-Device PII Scrubbing
5-Layer Global Architecture
Open Memory Playground
SOTA Memory Design

The 5-Layer Architecture

Inspired by MemGPT and human cognitive science. Agents process raw interaction in Working Memory, which is then distilled into specialized long-term storage layers by the background pipeline.

Explore Memory Dashboard

Working Memory

Active context

DISTILLING

Episodic Memory

Narrative history

Cold Archive

Cold storage log

LEARNING

Semantic Database

Factual knowledge

Procedural Rules

Learned playbooks

Hover to pause · Click a level to inspect

Hot Context
70% FULL
Capacity4096 tokens
Archival Trigger

The agent's active consciousness and short-term context window. Automatically archived when capacity threshold is reached.

Knowledge Ingestion

Enrich Memory at Machine Speed

Bridge the gap between static documents and active agent intelligence. Our multi-format pipeline structures unstructured data into reachable, high-fidelity knowledge.

Multi-Format Ingestion

Seamlessly import knowledge from JSON, YAML, or unstructured Raw Text documents.

Semantic Auto-Extraction

Transform raw sentences into structured semantic triplets using state-of-the-art LLM pipelines.

Automated Guardrails

Every fact is verified against organizational policies and scope levels before ingestion.

Real-Time Enrichment

Instantly broaden your agent's cognitive reach with zero-latency memory updates.

Studio_v2.0
Source: Raw Text

"The SLURM cluster at 10.50.0.0 uses H100 GPUs..."

Extracted Triplet
subject: "SLURM Cluster"
predicate: "EQUIPPED_WITH"
value: "NVIDIA H100 GPUs"
Access Control

Additive Scope Clearance

Multi-tenant AI requires strict data boundary enforcement. LearningX uses an elegant additive clearance model embedded at the vector database level — data never leaks between boundaries.

Enter Memory Playground
L1
L2
L3
L4
L5

Hover to pause · Click a level to inspect

L5Session

Session Scope

The most privileged scope. Session memory is locked to a single conversation ID and auto-expires. It captures in-flight context that should never persist beyond the session end.

Additive Access Rule

Only agents with exact L5 (Session) clearance can read this.

Visible Data Scopes
Global Scope
Organization Scope
User Scope
Agent Scope
Session Scope
Extensible Plugins

Bring Your Own Physics

The 5-layer architecture is driven by a chain of middlewares known as "Memory Policies." Want a custom embedding model? Strict PII redaction? Just inject your policy into the chain.

Configure Policy Chains
Limit by Size

Archives working memory when the token threshold is exceeded, triggering the persistence pipeline.

Adaptive Retention

Intelligently selects what to archive vs. permanently discard when working memory is consolidated.

Data Privacy Scrubbercustom

Strips personally identifiable information before memories are persisted. Bring your own regex or ML classifier.

Relevance Filter

Scores each candidate memory for relevance and drops anything too low-signal for long-term storage.

LLM Fact Arbitrator

Detects and resolves contradictions when new facts conflict with existing stored memories.

Exponential Forgetting

Applies memory decay — older, infrequently-accessed memories fade naturally, keeping storage lean.

Your Custom Policy

plugins.your_org.custom_policy

Hover to pause · Click a policy to inspect

native

Limit by Size

Archives working memory when the token threshold is exceeded, triggering the persistence pipeline.

CategoryArchival Trigger

Key Behaviors

Token budgetWorking memoryAuto-archival

Extend or Override

Native policies can be subclassed or swapped at any position in the chain. Your custom policy just needs to implement BasePolicy.

Continuous Learning

Memory Evolution & Decay

Not all knowledge is equal. LearningX evaluates every piece of information to determine its saliency. Crucial facts are reinforced, while outdated context naturally decays over time.

Manage Evolution Pipeline
0.0
1.0

Hover to pause · Click a point to inspect

Current Lifecycle

Raw Input

Unstructured conversational turns and logs.

Saliency Score0.2
Enterprise Grade

Built for Production Reality.

Standard RAG breaks down at scale. False facts compound, PII leaks into embeddings, and generic extraction misses industry nuance. LearningX solves the hard problems of AI memory.

Bi-temporal Knowledge

Facts are never deleted. When the truth changes, the old fact is retired — not overwritten. Agents always query the current state while compliance systems can rewind to any point in time.

Audit trail
Conflict-safe
Zero data loss

Local PII Scrubbing

Extract wisdom, not liabilities. Sensitive personal data is detected and redacted on-device before any memory is persisted — never sent to an external service or stored in any vector.

Runs locally
No egress
GDPR-ready

Domain Ontologies

Tailor the extraction lens to your industry. Load high-fidelity ontologies so the engine knows exactly which concepts, relationships, and entities matter in your domain — and ignores the noise.

DevOps
Healthcare
Finance
Hierarchical Configuration

Global Rules. Personal Touches.

Set company-wide memory rules once, then tailor how each AI agent thinks and remembers. No engineering team needed to change how an agent behaves.

Manage Agent Policies
Step 01

The baseline every agent inherits

Global Defaults

Set organisation-wide rules for how long memories live and how quickly the system forgets. Every agent starts here before you personalise.

Data retentionComplianceDecay speed
Persona
Step 02

Every agent thinks differently

Agent Personas

Override the global defaults per agent type (e.g. SRE vs Support). Tailor short-term focus or long-term recall based on the agent mission.

Memory depthFlush speedPersona bias
Session
Step 03

Dynamic context override

Chain Config

The ultimate override. Modify policy parameters per execution chain or specific user session for ultra-fine-grained control.

Veto thresholdsChain custom-TTLContext injection
Full Observability

Inside the Black Box.

Stop guessing what your AI knows or how it formed a belief. Every memory has a full audit trail — what triggered it, how confident the system was, and whether it's still relevant today.

  • See what your AI learned

    Browse every memory your agent formed, when it was created, and why it was kept in clear English.

  • Track memory health over time

    Monitor saliency scores, decay curves, and see which memories are fading or growing over time.

  • Audit every decision

    Full trace of what was extracted, what was rejected, and why — so you can trust what your AI knows.

Open the Memory Playground
Memory Trace — Session #4821
Conversation received
0.1s
Relevance scored
0.4s
Memory saved
0.8s
Linked to user profile
1 memory saved · 0 rejectedComplete
SOTA Monitoring

observe the
Subconscious.

Five levels of granularity to monitor your agents' health. From high-level system pulses to exact forensic lineage traces.

Forensic Deep-Dive
TRACE_#8192
ECL Source context
User: I'm planning to move to London next month.
Assistant: That's exciting! I'll update your preferences.
Confidence98.4%
Latency142ms
Live Pipeline

Trusted by developers at DecisionX to maintain
high-fidelity autonomous memories.

View Monitoring Dashboard
Ready for Production. Free to start.

Give your agents a mind.

Take control of AI memory isolation, behaviour tuning, and long-term knowledge — all in one platform built for teams who ship.