Skip to content

tkalevra/FaultLine

Repository files navigation

FaultLine Logo

FaultLine

FaultLine is a self-correcting memory system for LLM applications.

Most AI memory systems degrade over time — accumulating stale, incorrect, and conflicting information.

FaultLine does the opposite.

It enforces a controlled write path, builds structured knowledge from interaction, and continuously converges toward higher-quality memory through validation, reinforcement, correction, and decay.


Why FaultLine Exists

LLM memory systems fail in production for predictable reasons:

  • They trust the model to write correct data
  • They store everything without validation
  • They accumulate stale or contradictory facts
  • They have no mechanism for correction or decay

Over time, this leads to:

  • retrieval noise
  • conflicting context
  • degraded agent performance

FaultLine solves this by treating the LLM as an untrusted writer.

Memory is not stored by default — it is evaluated, structured, and earned.


What It Does

FaultLine transforms raw interaction into structured, governed memory:

  • Extracts entities and relationships from conversations
  • Infers missing structure (hierarchies, causality, relationships)
  • Validates all writes against a dynamic ontology
  • Builds a continuously evolving knowledge graph
  • Promotes high-confidence knowledge to long-term memory
  • Removes unused or unreinforced knowledge automatically
  • Allows real-time correction of facts and structure

Core Mechanics

Write Validation

No information is persisted without passing a validation gate.

Memory Promotion (C → B → A)

Facts move through a lifecycle:

  • Class C — Candidate (ephemeral, low confidence)
  • Class B — Confirmed (reinforced through interaction)
  • Class A — Canonical (user-confirmed truth)

Reinforcement

Frequently referenced knowledge strengthens and persists.

Decay

Unused or unreinforced knowledge expires automatically.

Structural Inference

FaultLine builds missing relationships:

  • hierarchy (family → members → pets → animals)
  • causality
  • relational links

Live Correction

Users can modify the graph through natural interaction:

"my name isn’t Todd, it’s Bradly"
"my pets are not part of my family"

FaultLine updates:

  • facts
  • relationships
  • graph structure

All changes are non-destructive and auditable.


How It’s Different

Traditional Systems FaultLine
Trust model output blindly Treat model as untrusted writer
Store everything Validate before persistence
Accumulate noise over time Decay unused knowledge
Static or no schema Self-evolving ontology
Hard to correct Interactive correction
Requires cleanup Self-maintaining

System Model

Agent Runtime (OpenWebUI / LangChain / etc.)

FaultLine (extract → validate → infer → promote → decay)

PostgreSQL (graph) + Qdrant (semantic)

FaultLine sits beneath your application.

  • Your agent writes → FaultLine decides what becomes memory
  • Your agent reads → FaultLine returns structured knowledge

Retrieval Model

FaultLine uses multi-path retrieval:

  • Graph traversal (relationships, hierarchy)
  • Semantic similarity (Qdrant)
  • Confidence-weighted facts (A/B/C classes)

This ensures:

  • weak signals remain visible
  • strong knowledge dominates
  • incorrect structure can be corrected through use

Design Philosophy

Memory must be earned

Inference alone is not enough — knowledge must be used and reinforced.

Memory must adapt

Facts, relationships, and structure must be correctable through interaction.

Memory must decay

Unverified or unused knowledge should not persist indefinitely.


Positioning

FaultLine is not:

  • a vector database
  • a RAG system
  • an agent framework

It is:

→ a governed memory layer that ensures long-term memory remains usable

Use it with:

  • OpenWebUI
  • LangChain
  • custom agents

Quick Start

git clone https://github.com/your-org/FaultLine.git
cd FaultLine

cp .env.example .env
# edit .env

docker compose up -d

Development Status

  • Core architecture implemented ✅
  • Self-correcting lifecycle ✅
  • Dynamic ontology ✅
  • Production-scale validation → in progress

License

Licensed under the MIT License — see LICENSE

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Packages

 
 
 

Contributors

Languages