Asset Core: Spatial Memory for AI Agents

Why Agents Hallucinate Spatial State

LLM-based agents fail at multi-step spatial tasks because they lack authoritative memory for “where things are.” When agents query environment state, plan actions, and execute tool calls across multiple turns, they encounter:

• State drift: Agents cache stale positions; world state diverges from agent beliefs
• Non-atomic updates: Partial operation failures leave corrupted spatial configurations
• Replay failures: Debugging requires manual log forensics; exact state reconstruction is impossible
• Coordination collisions: Multiple agents overwrite each other’s spatial changes
• Audit gaps: No provable record of “who moved what, when, and why”

These aren’t LLM reasoning failures—they’re infrastructure failures.

The model is stateless. The world must be stateful. The model is probabilistic. The world must be deterministic.

Asset Core provides that world.

Authoritative Spatial Memory for Stateless Agents

Asset Core is a deterministic spatial-transactional engine that serves as the ground-truth memory layer for LLM-based agents operating in spatial environments.

Every state change is sealed by a single append to the durable commit log; all queries and notifications are derived from that one source of truth, eliminating dual-write drift. This makes Asset Core the authoritative spatial memory for stateless agents where correctness, determinism, and proof matter as much as performance.

For agent platforms

• Authoritative state queries: Agents query current world state efficiently
• Transactional tool calls: Multi-step operations commit atomically or abort completely
• Deterministic replay: Every state transition is reproducible from the commit log
• Multi-agent coordination: Global commit ordering prevents conflicting spatial updates
• Audit-grade provenance: Full causal history—which agent, which tool call, which context

For researchers

• Reproducible experiments: Identical event sequences guarantee identical outcomes
• Spatial semantics: First-class support for 0D (aggregates), discrete grids (1D/2D), and continuous lines/planes (1D/2D); 3D spaces on the roadmap
• Universal operations: add/remove, move, split/merge, transfer work consistently across all container types
• Event sourcing: Hybrid events (delta + post-state) enable both real-time analytics and perfect replay
• Performance at scale: Structure-of-Arrays layout, dense IDs, cache-friendly iteration for high-throughput workloads

What Agents Can Do Today

Discrete spatial containers:

• 0D containers (Balances, Slots): For agents managing resources, equipment, inventories without spatial geometry
• 1D grids (ℤ): For agents reasoning about linear sequences, conveyors, ordered lanes
• 2D grids (ℤ²): For agents navigating warehouses, game boards, floor plans with collision detection

Continuous spatial containers:

• 1D continuous lines (ℝ): For rails, actuators, and single-axis robotics with fixed-point precision
• 2D continuous planes (ℝ²): For robot workcells, pick-and-place, and metric collision checks

Universal operations:

• add / remove: Create or destroy entities in containers
• move: Relocate entities within or across containers
• split / merge: Partition resources or combine them
• transfer: Cross-container state transitions with consistent semantics

Agent-friendly guarantees:

• Same events ⇒ same state (deterministic replay)
• Operations commit atomically or abort completely (no partial failures)
• Preflight validation for non-mutating commit checks with commit-equivalent errors
• Multi-tenant isolation (each agent session gets independent world state)
• Audit-grade logs (every state change recorded with full provenance)

On the roadmap:

• 3D discrete grids (ℤ³) for volumetric reasoning
• Continuous 3D spaces (ℝ³) with physics integration

Who It’s For

Asset Core is designed for teams building:

• AI labs and agent platforms: Building LLM-based agents that plan, act, and coordinate across spatial environments—providing deterministic world state for tool-calling agents, multi-agent coordination, reproducible evaluation, and audit-grade compliance
• Multi-agent researchers: Investigating agent coordination, spatial reasoning, and reproducible multi-step workflows with shared world models and deterministic debugging
• Simulations & games: Inventories, spatial logic, and deterministic save/replay for virtual environments

How to Access

Asset Core is proprietary software distributed as binary-only under specific licenses:

Learn More

For technical details, see Basics for architecture and container types, Examples for concrete container demonstrations, or the documentation for complete API specifications.