Features
Outcome-based memory scoring (wilson score)
Two-lane retrieval: 4 summaries + 4 facts per turn
TagCascade retrieval pipeline
Automatic summarization of conversations
Noun tagging at store time
Memory lifecycle: working (24h), history (30d), patterns (∞)
Demotion of low-scoring memories
Poison-resilient against semantic match + spoofed trust
Works with Claude Code and OpenCode
Local-first, no cloud dependency
Desktop GUI with Ollama/LM Studio support
MCP (Model Context Protocol) tools
Trained exclusively on production code for maintainability
Supports COBOL, Fortran, Verilog, Rust, SQL
Terminal-native CLI with local-to-remote execution
External tool access via MCP
Multi-agent orchestration via Cosine Swarm
Cosine Cloud for parallel tasks and shared projects
Scout model: on-device, post-trained from Devstral 123B
Outpost model: 59.3% on Niche-Bench, post-trained from Kimi K2.6
Frontier model: coming soon for frontier-scale reasoning
Air-gapped deployment option
Managed single-tenant cloud
Credit-based usage metering
CLI built-in search (v1.16)
Lumen Sovereign coalition for sovereign AI
