Most agents are brilliant.Almost none of them remember.
Nous is a continuous memory platform for AI agents. It gives any LLM persistent memory, decision intelligence, and the ability to learn from its own mistakes — an architecture that treats memory as part of cognition rather than as a database bolted to the side.
- Automated tests
- 5,573
- Feature specs
- 75
- Commits
- 1,151
across 303 files
written before the code
single-engineer project
Figures generated from the Nous source repository, 2026-08-11.
The problem
Statelessness is not a missing feature. It is the ceiling.
An agent that forgets cannot compound. It repeats corrected mistakes, rebuilds context the user already supplied, and has no way to tell a judgement that worked from one that merely sounded convincing.
Stateless agent
- Repeats mistakes it has already been corrected on
- Rediscovers preferences every session
- Accumulates no skill from work it has already done
- Cannot say whether its confidence was ever justified
With Nous
- Corrections become censors — a mistake blocks itself next time
- Five memory types, each with a distinct retention rule
- Procedures are learned, scored, and reused across sessions
- Brier scoring measures whether confidence tracks reality
Architecture
Brain and Heart
Following Minsky's Society of Mind, cognition is split into two organs joined by an event bus: one that judges, one that remembers. Neither is useful alone.
Brain
Evaluation and judgement- Quality scoring
- LLM-judged response quality against a multi-dimensional rubric.
- Brier calibration
- Tracks whether stated confidence matches real outcomes.
- Decision tracking
- Every decision logged with its outcome, so history is evidence.
- Knowledge graph
- Relationships across facts, entities and episodes via spreading activation.
- CEL guardrails
- Common Expression Language rules enforcing behavioural bounds.
Heart
Memory and identity- Episodes
- Conversation summaries, temporally indexed and importance-scored.
- Facts
- Extracted knowledge with confidence, categories and staleness decay.
- Procedures
- Learned skills with triggers, tools and effectiveness ratings.
- Working memory
- Per-session context: active frame, loaded facts, execution ledger.
- Censors
- Guardrails that block known-bad patterns before they execute.
Memory architecture
Five kinds of memory
Modelled on human memory systems. Each type has its own write path, retention rule and retrieval strategy — collapsing them into one store is what makes most memory layers degrade as they grow.
Episodic
What happened and when. Conversation summaries with temporal ordering and importance scoring.
Semantic
What the agent knows. Facts carrying confidence scores, categories and staleness decay.
Procedural
How the agent works. Skills with triggers and effectiveness ratings, evolved via EvoSkill.
Working
What is live right now. Current frame, loaded facts, active task, execution ledger.
Censors
What must not happen. Regex and semantic guardrails evaluated before execution.
Capabilities
What it actually does
Six of the 75 shipped feature specs. Each was written before its implementation, which is why they can be counted.
Sleep consolidation
UniqueA 16-phase autonomous cycle — pruning, compaction, contradiction resolution, procedure learning, graph densification and more — that runs between sessions.
Correction learning
F039User corrections are detected and dual-written to facts and censors, so a mistake becomes a permanent guardrail.
EvoSkill
UniqueProcedural memory that proposes, tests and merges its own skills rather than waiting to be taught.
Decision intelligence
CoreConfidence, outcome and Brier score on every decision — the agent calibrates its own judgement.
Hybrid retrieval
CoreVector, keyword and graph expansion fused by Reciprocal Rank Fusion, then re-ranked for diversity.
Execution integrity
CoreAction gating, claim verification and an execution ledger — the defence against hallucinated actions.
Foundations
Built on published work
Every architectural choice traces to a paper rather than to taste. The research below is what the design is accountable to.
Minsky
1986Society of Mind — organ duality, frames, censors, K-lines
MemAlign
2025Correction learning — dual-memory mistake capture
A-MEM
2025Agentic memory — self-organising knowledge evolution
TIM
2025Trajectory learning — improving from execution paths
ACC
2025Adaptive context control — dynamic memory management
xMemory
2025Diversity-aware retrieval — MMR re-ranking
- 77,369
- Lines in the platform package
- 150
- Modules
- 62
- Schema migrations
- 5,573
- Tests
Memory is the part that compounds.
The architecture, the feature specs and the research trail are all public. Start with the docs.