Design Principles (Plain English)

FaultLine is built to play nicely with every layer it touches. Nothing is forced, hacked, or "almost compatible".

How FaultLine Compares

Most AI memory systems trust the LLM to write whatever it extracts. FaultLine doesn't — every fact passes a validation gate before it touches storage. It's the only system in the field that treats the model as an untrusted writer by design.

Project	Open Source	Self-Host	Write Gate	Knowledge Graph	Short → Long Term	Per-User	Temporal	OpenWebUI
FaultLine	✅ MIT	✅	✅	✅	✅ A→B→C pipeline	✅	✅	✅
Graphiti (Zep)	✅ Apache	✅	❌	✅	❌	⚠️	✅	❌
Zep Cloud	❌	❌	❌	✅	❌	✅	✅	❌
Mem0	⚠️ open-core	⚠️	❌	⚠️ paid	❌	✅	⚠️	❌
Letta / MemGPT	✅ Apache	✅	❌	❌	⚠️ agent-managed	✅	❌	❌
Cognee	✅ Apache	✅	❌	✅	❌	⚠️	⚠️	❌
EverMind / EverOS	✅ Apache	✅	❌	✅	⚠️ expiry only	⚠️	✅	❌
LangMem	✅ MIT	✅	❌	❌	❌	⚠️	❌	❌
LlamaIndex Memory	✅ MIT	✅	❌	⚠️	❌	⚠️	❌	❌
LightRAG	✅ MIT	✅	❌	✅	❌	❌	❌	❌
SuperLocalMemory	✅ Apache	✅	❌	❌	❌	❌	❌	❌
OMEGA	✅ Apache	✅	❌	❌	❌	❌	❌	❌
Memary	✅ MIT	✅	❌	✅ Neo4j	❌	❌	❌	❌
Motorhead	✅ MIT	✅	❌	❌	❌	⚠️	❌	❌
CrewAI Memory	✅ MIT	✅	❌	❌	❌	❌	❌	❌
Hermes Agent	✅ Apache	✅	❌	⚠️	⚠️ episodic	❌	❌	❌
Hindsight	❌	❌	❌	⚠️	❌	✅	⚠️	❌
Supermemory	❌	⚠️ enterprise	❌	⚠️	❌	✅	❌	❌
Cloudflare Agent Memory	❌	❌	❌	❌	❌	✅	❌	❌
Microsoft MAF	✅ MIT	⚠️ Azure	❌	❌	❌	⚠️	❌	❌

✅ Full · ⚠️ Partial or paywalled · ❌ Not supported

Short → Long Term refers to staged fact promotion: facts move from ephemeral (Class C) through behavioral (Class B) to permanent storage (Class A) based on confirmation, rather than writing everything to long-term storage immediately or never at all.

FaultLine is the only entry with a validated promotion pipeline — unconfirmed facts expire, confirmed facts persist, and nothing is written unsupervised at any stage.

---

1. Developer‑compliant by design

Each component is used the way its authors intended:

• GLiNER2

  • Receives schema‑typed inputs, not raw prompts
  • Outputs structured entities and relations exactly in its preferred format
  • No prompt abuse, no post‑hoc guessing

• OpenWebUI

  • Uses official inlet filters and function hooks
  • Context is injected using supported message roles
  • Nothing bypasses or mutates internal state unexpectedly

• PostgreSQL

  • Single‑transaction writes
  • Explicit uniqueness constraints prevent duplication
  • Facts are stored as data, not blobs of text

• Qdrant

  • Used strictly as a derived index
  • Never treated as source‑of‑truth memory
  • Rebuilt safely from Postgres when needed

Result:
If any component is swapped, upgraded, or audited, FaultLine remains predictable and maintainable.

---

2. Memory is strictly separated per user

• Each user has isolated long‑term memory
• No facts are shared across users
• No embeddings are cross‑user
• Queries are always scoped by user_id

This is not just privacy‑friendly — it's correct.

"The assistant remembers me" does not mean
"The system remembers everyone"

---

3. Short‑term vs Long‑term memory (human‑style)

FaultLine models memory the way people do it:

• Short‑term memory
  What's being talked about right now (the current conversation)

• Long‑term memory
  Verified facts that persist across conversations

• Fast recall
  "This seems related" hints used to surface helpful context

Nothing automatically becomes long‑term memory.
It must pass validation first.

---

Why This Improves Answers (Concrete Example)

Without FaultLine (typical AI behavior)

User:

"What's the weather like tomorrow?"hing automatically becomes long‑term memory.
It must pass validation first.

---

Why This Improves Answers (Concrete Example)

Without FaultLine (typical AI behavior)

User:

“What’s the weather like tomorrow?”

Assistant:

“I don’t know where you are.”

Why this happens:

• Location was mentioned earlier
• It fell out of the conversation window
• It was never stored as a real fact

---

With FaultLine

User:

“What’s the weather like tomorrow?”

System already knows:

• User lives in Mianus
• That fact was validated and stored earlier

Assistant:

“Tomorrow in Mianus, expect…”

No guessing. No asking again. No hallucination.

---

Why This Is Different From Vector‑Only Memory

Most “memory systems” today:

• Store chunks of text
• Search by similarity
• Hope the retrieved text is correct

FaultLine:

• Stores validated facts
• Checks for conflicts
• Uses vectors only as a hint system, never as truth

Vectors help you find memory
Facts decide what memory is true

---

Extremely Simple Mental Model

Flow in plain words:

1. User says something
2. It lives in short‑term memory
3. If it looks like a real fact, it’s checked
4. If valid, it’s written to long‑term memory
5. Related memories are recalled to improve answers

---

Why This Matters

FaultLine doesn’t try to be clever — it tries to be correct.
It respects component boundaries, keeps user data isolated, separates conversation from memory, and only remembers things that pass validation.

The result is an assistant that feels more human:

• It remembers what matters
• It forgets what doesn’t
• It doesn’t ask the same questions over and over

---

Novel concepts drawn from

#	Novel Aspect	Status in the Wild	Reference
1	Write Gate	Theorized in research, never shipped in production	arxiv 2603.15994
2	Fact Lifecycle	Benchmarks confirm most systems fail on selective forgetting — no production system implements classified promotion	arxiv 2603.07670
3	Self-Building Ontology	Open research problem; all existing approaches require static schemas	arxiv 2604.20795
4	Mnemonic Sovereignty	Framed as a normative goal in 2026 security literature — no deployed system implements it end-to-end	arxiv 2604.16548
5	Metadata-Driven Validation	Memory governance decoupled from evolution identified as an unsolved problem	arxiv 2603.11768