Reasoning
Plans like a strategist. Grounded like a lawyer.
The Hierarchical Reasoning Module plans abstractly, executes concretely, and decides when to stop. Neurosymbolic fusion grounds every embedding in a symbolic path — hallucinations become a hardware problem you can fix.
Capabilities
What it does.
Every capability is production-grade. No flags, no betas — these ship the day you adopt Minds.
H-Module · abstract planner
Decomposes the goal, orders subgoals, decides when ACT stops.
L-Module · concrete executor
Tool calls, KG lookups, vector recall, chain-of-thought.
Adaptive Computation Time
The model decides when it has thought enough.
Neurosymbolic fusion
Neural similarity blended with symbolic graph paths per query.
Microtheory scoping
Reason inside a worldview — finance, medical, legal — without crosstalk.
Causal discovery
BFS, Dijkstra, A* path-finding over the knowledge graph.
API
A few lines is all it takes.
The SDK reflects the conceptual model directly. No glue code, no orchestrator, no learning curve beyond the data shape.
exampletypescript
const plan = await minds.reason({
goal: "Why did Q3 revenue dip?",
mode: "neurosymbolic",
blend: 0.6, // 0=pure symbolic, 1=pure neural
microtheory: "finance",
maxDepth: 6, // HRM ACT cap
});
console.log(plan.steps); // hierarchical plan
console.log(plan.confidence); // 0.94
console.log(plan.path); // KG path: revenue → channel-mix → ...Specs
What it costs, what it guarantees.
Performance
- Backend
- Burn (Rust ML)
- Models
- NeoBERT 250M + Qwen3-0.6B (~2.5 GB)
- Plan latency
- <420ms for depth-6 plans
- Path-finding
- BFS · Dijkstra · A*
- Counterfactuals
- first-class API
Guarantees
- Durability
- fsync per WAL commit
- Isolation
- MVCC · snapshot
- Audit
- BLAKE3 Merkle chain
- Encryption
- AES-256-GCM per-namespace
- Concurrency
- 100K+ ops/sec
Build it on Minds.
Start with the SDK. Ship in an afternoon. Run anywhere — Akasha Cloud, on-prem, or air-gapped.