AiDENs

Overview

AiDENs is a closed-loop self-learning AI system built in Rust. It runs autonomously without human intervention, introspecting its own typed knowledge graph to find structural and content-level gaps, generating prioritized tasks to fill them, executing those tasks via a local LLM (Ollama), capturing the results as provenance-attributed facts with graph edges, evaluating fact quality through a governance gate, and recording RL routing feedback for adaptive retrieval.

The system combines 56K LOC across 36 crates with 668 tests (506 core + 162 autonomous). It runs entirely local-first — the only external cost is the LLM model (local Ollama or cloud models proxied through it).

The Closed Loop

Each cycle of the autonomous loop:

1. Detect — Scans the knowledge graph for gaps via 8 mission types (contradictions, stale facts, missing context, duplicates, file reference verification, codebase sync, provenance tracing, namespace completeness)
2. Generate — Converts detected gaps into prioritized JobV1 entries with idempotency keys, enqueued to a durable queue with leases
3. Execute — Dequeues jobs, builds a PlanActVerifyLoopV1 with persistent memory, sends the prompt to the LLM via Ollama
4. Capture — Extracts individual factual statements from the model's response, adds each as a new fact to semantic-memory with source attribution and graph edges
5. Evaluate — Scores each captured fact on content quality (numbers, dates, proper nouns, technical terms, coherence) — promotes high-quality facts, quarantines medium, rejects low
6. Feedback — Records RL routing outcome via /record-outcome for adaptive retrieval learning

Mission System

Eight high-ROI mission types with adaptive priority scheduling. Missions that find 0 issues get lowered priority; missions that find issues reset to base priority.

• VerifyPublishedCrates (0.9) — Check if facts about crate versions match current state
• DetectContradictions (0.85) — Find pairs of facts with conflicting claims
• VerifyFileReferences (0.8) — Check if file paths mentioned in facts still exist
• VerifyCodebaseSync (0.75) — Check if facts about codebase metrics are still accurate
• TraceProvenanceChains (0.7) — Verify source attributions still exist and support claims
• StaleDateDetection (0.65) — Find facts referencing dates more than 6 months old
• FindDuplicates (0.6) — Find facts with high content overlap across namespaces
• AuditNamespaceCompleteness (0.5) — Check each namespace for graph connectivity

What Doesn't Exist Elsewhere

No single system, and no combination of systems, provides this combination. The closest neighbors:

• Zep/Graphiti — Has a temporal knowledge graph but is a passive memory backend (no autonomous loop, no gap detection, no provenance, no evaluation gate)
• Letta/MemGPT — Has self-editing memory but reactive (during conversation), not autonomous. No typed graph, no gap detection, no evaluation gate
• AutoGPT — Has autonomous loops but creates tasks from user goals, not from KB introspection. No persistent knowledge graph, no gap detection, no provenance
• MemRL — Academic concept for RL memory evolution. Not implemented. Much narrower scope

The novelty is in the combination: autonomous KB gap detection + typed knowledge graph with provenance + closed loop with evaluation gate + mission system + RL routing feedback + durable queue with safe mode + local-first operation.

Technical Architecture

• semantic-memory — The knowledge graph: 7000+ facts, typed edges (semantic/temporal/causal/entity), bitemporal search, contradiction detection, factor graph belief propagation, 48 MCP tools, 15 HTTP endpoints, RL routing
• aidens-autonomous — The closed-loop brain: gap detector (7 gap types), mission system (8 mission types), executor (PlanActVerifyLoopV1 with Ollama), capture (sentence-level extraction), evaluation (content quality scoring), loop driver (safe mode, attempted gaps tracking)
• aidens-tui — Terminal UI with 4 panels: memory stats, loop state, queue, activity log

Provenance First

Every fact the system learns has source attribution, confidence, and a receipt. The system is bitemporal — it knows what it believed on date X and can trace how its beliefs changed. It can supersede its own beliefs when it finds better evidence. It checks its own DB health before acting.

Live Demo

A live demo page shows the system in action with real metrics, captured facts, and the comparison table against other systems.