AI That Remembers, Learns, and Gets Smarter
Nous is a continuous memory platform for AI agents — inspired by Minsky's Society of Mind and grounded in cognitive science. It gives any LLM persistent memory, decision intelligence, and the ability to learn from its own mistakes.
AI Agents Are Goldfish with PhDs
Every major AI agent today forgets everything between sessions. They're brilliant — and amnesiac.
Without Continuous Memory
Today's AI agents are stateless by default. Every conversation starts from zero.
- Repeated mistakes — no correction learning
- Lost context — user re-explains preferences every session
- No skill accumulation — can't get better at what it does
- No decision tracking — no idea what worked or failed
- No identity — each session is a different agent
With Nous
A cognitive memory layer that makes any LLM agent persistent, learning, and self-improving.
- Correction Learning — mistakes become permanent guardrails
- 5 memory types — episodic, semantic, procedural, working, censors
- Sleep consolidation — autonomous overnight memory optimization
- Decision intelligence — Brier-scored prediction calibration
- Persistent identity — same agent personality across all sessions
Brain / Heart Duality
Inspired by Minsky's Society of Mind — cognition split into evaluation (Brain) and memory (Heart)
🧠 Brain
- ⚡ Quality Scoring — LLM-judged response quality with multi-dimensional rubric
- 🎯 Brier Calibration — tracks prediction accuracy over time, self-correcting confidence
- 📊 Decision Tracking — every decision logged with outcome, enabling learning from history
- 🕸️ Knowledge Graph — relationships between facts, entities, episodes via pgvector + spreading activation
- 🔗 Bridge Definitions — cross-concept connections that enrich retrieval
- 🛡️ CEL Guardrails — Common Expression Language rules enforcing behavioral bounds
❤️ Heart
- 📖 Episodes — conversation summaries with emotional valence, temporally indexed
- 💡 Facts — extracted knowledge with confidence scores, categories, decay tracking
- ⚙️ Procedures — learned skills with triggers, tools, and effectiveness ratings
- 🔄 Working Memory — dynamic per-session context: frame, task, loaded facts, recent episodes
- 🚫 Censors — behavioral guardrails that block harmful patterns before they execute
- 🪪 Identity — persistent personality and initiation protocol across all sessions
The 7-Phase Cognitive Loop
Every interaction passes through a biologically-inspired processing pipeline
Five Memory Types
Modeled after human cognitive memory systems — each type serves a distinct purpose
Episodic
Conversation summaries with emotional valence, temporal ordering, and importance scoring. What happened and when.
Semantic
Extracted facts with confidence scores, categories, tags, and staleness decay. What the agent knows.
Procedural
Learned skills with triggers, tool lists, instructions, and effectiveness ratings. Self-evolving via EvoSkill.
Working
Per-session dynamic context — current frame, loaded facts, recent episodes, active task, execution ledger.
Censors
Behavioral guardrails that block harmful patterns before execution. Regex + semantic matching with severity levels.
39 Shipped Features
Production-hardened capabilities that no other memory platform offers
Sleep Consolidation
5-phase autonomous cycle: compaction, fact extraction, knowledge graph enrichment, decision calibration, amnesia prevention. Runs between sessions.
UniqueCorrection Learning (F039)
Dual-path detection of user corrections → dual-write to facts + censors. Mistakes become permanent guardrails. MemAlign-inspired.
LiveEvoSkill
Self-evolving procedural memory. Skills are proposed, tested, merged, and improved autonomously. Transferable and composable.
UniqueDecision Intelligence
Every decision tracked with confidence, outcome, Brier scoring. The agent literally calibrates its own judgment over time.
UniqueKnowledge Graph
Spreading activation across entities and relationships. Graph-expanded retrieval surfaces connections vector search alone would miss.
ShippedHybrid Retrieval
Vector + keyword + graph expansion, fused via Reciprocal Rank Fusion (RRF), then MMR diversity re-ranking. Research-grade retrieval.
ShippedCognitive Frames
Dynamic context modes (task, research, creative, etc.) that reshape working memory, tool access, and retrieval strategy per situation.
ShippedDAG Orchestration
Multi-step workflow pipelines with dependency tracking, 3-state exit codes, wave-based execution, and dashboard monitoring.
LiveExecution Integrity
Action gating, claim verification, execution ledger, intent tracking. Prevents hallucinated actions and duplicate operations.
ShippedWhat Makes Nous Different
Head-to-head advantages vs. Mem0, Letta, MemOS, and stateless agents
Cognitive Architecture, Not Just Storage
Competitors store memories. Nous thinks with them — cognitive loop, frames, deliberation, monitoring. Memory is part of cognition, not bolted on.
Self-Improving Intelligence
Brier scoring, correction learning, EvoSkill, sleep consolidation. Nous gets measurably smarter with every interaction — competitors don't.
Research-Grounded Design
Built on 12+ cognitive science and AI research papers (Minsky, MemAlign, A-MEM, TIM, ACC). Every feature traces to published research.
Production-Grade Safety
Censors, action gating, claim verification, CEL guardrails, execution ledger. Enterprise-ready safety that memory-only platforms lack entirely.
Brain/Heart Duality
Unique organ architecture separating evaluation from memory. No other agent platform has this — it enables independent scaling and specialization.
4 Unique Capabilities
Pre-prune fact extraction, usage tracking feedback, anti-hallucination safety, and SmartCompress type-aware compression. Verified unique in competitive analysis.
Built to Scale
Standing on Giants
Every architectural decision traces to published cognitive science and AI research
Minsky (1986)
Society of Mind — organ duality, frames, censors, K-lines
MemAlign (2025)
Correction learning — dual-memory mistake capture
A-MEM (2025)
Agentic memory — self-organizing knowledge evolution
TIM (2025)
Trajectory learning — improving from execution paths
ACC (2025)
Adaptive context control — dynamic memory management
tinyHippo
Biological memory model — hippocampal consolidation
Membrain
Engineering memory taxonomy — type classification
xMemory (2025)
Diversity-aware retrieval — MMR re-ranking