CoreTx

Your AI remembers your name and your preferences. It does not remember the architectural constraint you spent 40 minutes establishing last Tuesday. It hallucinates facts you told it explicitly. It gets worse over time as stale information piles up. And switching between Claude and GPT means starting your knowledge from zero.

The memory features that exist today are flat text: a list of things the model was told, with no sense of what matters, what changed, what contradicts what, or how your understanding evolved. That is not how brains work, and it is not how knowledge should work.

CoreTx adds a brain-inspired knowledge layer to your AI. It extracts structured claims from your conversations, scores them for importance before storing, tracks how your understanding evolves over time, detects contradictions, and lets old noise fade while reinforcing what you actually use. It works across every AI provider you use. Your knowledge belongs to you.

CoreTx implements both halves of how biological memory works. Your AI keeps a compressed narrative of past conversations (episodic memory) alongside structured facts extracted from those conversations (semantic memory). The narrative gives context for reasoning. The facts ensure specific decisions and constraints cannot drift or be hallucinated, even months later. Neither half is sufficient alone: summaries lose facts after compaction, and isolated facts lose the reasoning thread that connects them. Together they give you grounded reasoning, where the model can think through the story while anchored to facts that will not change.

This is the architecture described by Complementary Learning Systems theory (McClelland et al., 1995), which models the interplay between fast hippocampal encoding and slow neocortical consolidation. CoreTx's DREAMS consolidation plays the role of sleep replay, transferring fragile episodic memories into durable semantic knowledge. Other memory tools only have the summary half. That is why their accuracy degrades over time and ours does not.

The result is an AI that outperforms any frontier model on your accumulated work. Over months of use, it reasons over your constraints, your decisions, your evolved understanding with high accuracy, while flat text notes degrade and introduce errors. It connects knowledge across your projects: a bug you found in one codebase surfaces when you are about to make the same mistake in another.

Local-first architecture. Your data stays on your machine by default. Nothing is shared unless you choose to.

What changes

1. Remember what you decided last month. Context persists across sessions and context resets. Structured storage retains facts that flat text loses within a few compactions.
2. Get facts right about your own work. Facts are stored as typed, attributed knowledge objects. The system retrieves what you actually said, or tells you it doesn't know.
3. Use Claude, GPT, and Gemini on the same knowledge. Your knowledge is stored independently of any AI provider. Start a conversation in Claude Code, continue it in claude.ai, switch to GPT. Same knowledge, no re-explaining.
4. Turn off model training and still get smarter. The structured store learns from your usage patterns. Frequently accessed knowledge surfaces faster. Stale knowledge fades. No fine-tuning required.
5. Keep your project goals aligned. Standard memory loses most project goals after a few rounds of context compaction. Structured storage preserves them.
6. Filter noise before it enters your memory. A write gate scores every fact for novelty, reliability, and source quality before storing it. The store stays clean as it grows.
7. Let your memory self-organize. Brain-inspired architecture: episodic memories decay, semantic memories persist, related findings consolidate into higher-order insights. No manual curation needed.
8. Share selectively. Three privacy levels: private (only you), matchable (anonymous structural matching), public (visible to the network). You control what is shared, with whom, and when.

Why it works

Every component is modeled on how biological memory actually operates: salience filtering, forgetting curves, sleep consolidation, prediction error, associative recall. These are the mechanisms that let humans learn continuously without drowning in noise.

• Salience gating (hippocampal write gate)
  Not everything deserves to persist. A scoring function evaluates novelty, reliability, and specificity before admitting anything to the store.
• Vividness decay (Ebbinghaus forgetting curve)
  Episodic memories fade over time. Semantic memories persist. Accessed knowledge strengthens. The store self-cleans without manual curation.
• Consolidation (REM sleep replay)
  Related findings from different sessions cluster and merge into higher-order patterns. The system discovers connections you did not explicitly make.
• Contradiction detection (prediction error)
  When new information conflicts with stored knowledge, the system flags it rather than silently maintaining both versions.
• Version chains (memory updating)
  Understanding evolves. The system tracks how your knowledge changed over time, so it can answer both "what is true now" and "what did we believe then."
• Cryptographic provenance (attribution)
  When your AI builds on knowledge you created, you should get credit. Every piece of knowledge is cryptographically signed so authorship is always traceable.

How it works

Performance

Measured on published benchmarks. Full methodology in linked papers below.

Research

CoreTx is built on published research. Four peer-reviewed papers cover the core architecture.

Facts as First-Class Objects: Structured Knowledge Persistence for LLM Agents

Structured knowledge objects vs context windows. Compaction survival, multi-hop reasoning, cross-domain synthesis. arXiv:2603.17781

Salience-Gated Memory for Continual LLM Agents

Write-time filtering with novelty, reliability, and source reputation scoring. arXiv:2603.15994

ANML: Attribution-Native Machine Learning

Training methodology that preserves source attribution through the learning process. arXiv:2602.11690

Attention is Not Retention

Why transformer attention mechanisms fail at long-horizon knowledge retention. arXiv:2601.15313