Memory That Thinks.
Thermodynamic vMMU for AI Agents
Your agent forgets everything the moment its context window fills. SULCUS gives it a real memory — a thermodynamic graph that heats what matters, cools what doesn't, and pages the right context in at the right time. Token burn drops up to 90%. Recall goes to 100%.
Memories aren't static rows in a database. They're thermodynamic objects — born hot, cooling with time, reheating on recall, diffusing through edges to related knowledge. When an agent searches, the engine doesn't just find matches. It ignites the graph.
Three specialized vectors for perfect long-term recall.
The thermodynamic engine compiled to WebAssembly. Runs in your agent's process, in the browser, or on the edge. Sub-millisecond reads. Zero network calls. Your data never leaves the machine.
A native Rust process that sits between your agent and its LLM. It intercepts context, injects relevant memories, and pages out stale turns automatically. Works with any MCP-compatible host.
CRDT-based replication across agents, machines, and teams. Every agent maintains a local graph; the cloud merges them into a shared knowledge mesh. Conflict-free by design.
30+ Configurable Knobs. Zero Hardcoded Behavior.
Every parameter of the thermodynamic engine is exposed via API. Tune decay profiles, resonance depth, consolidation thresholds, and context budgets per tenant, per type, per node.
Reactive Memory Triggers
Set rules on your memory graph. When events happen — a memory is stored, recalled, boosted, or starts to decay — Sulcus fires actions automatically. Pin important memories. Boost fading context. Webhook your systems. The agent doesn't have to remember to remember.
Auto-pin every preference memory so it never decays.
Reinforce memories every time they're searched — spaced repetition, automated.
Alert the agent when critical knowledge starts cooling. Act before it's lost.
Fire HTTP callbacks when memory heat crosses a boundary. Integrate anything.
6 events · 7 actions · filter by type, namespace, label pattern, heat threshold · unlimited triggers per tenant
Five lines. Real memory.
Zero-dependency SDKs for Python and Node.js. Or connect directly via MCP or REST.
pip install sulcusfrom sulcus import Sulcus
s = Sulcus(api_key="sk-...")
s.remember("User prefers dark mode",
memory_type="preference")
results = s.search("dark mode")npm install sulcusimport { Sulcus } from "sulcus";
const s = new Sulcus({ apiKey: "sk-..." });
await s.remember("User prefers dark mode",
{ memoryType: "preference" });
const results = await s.search("dark mode");Your memories. Your machine. Your rules.
Most memory systems require your data to live on someone else's server. Sulcus runs locally by default. Cloud sync is optional, encrypted, and you control what leaves the machine.
The WASM core and MCP sidecar run entirely on your hardware. No network calls required for reads or writes. Your agent's memory never touches a cloud server unless you explicitly enable sync.
When you do enable cloud sync, data is encrypted in transit (TLS 1.3) and at rest (AES-256). Tenant isolation ensures your memory graph is invisible to other users — and to us.
Self-host the entire stack — server, database, sync — in your own infrastructure. The MIT-licensed core means no vendor lock-in, no phone-home telemetry, no surprises.
Choose which namespaces sync to the cloud and which stay local. Sensitive memories (credentials, personal data, health info) can remain on-device while shared knowledge replicates across your fleet.
Free tier. No credit card. Start building agents with real memory in under five minutes.
Privacy-first. Secure. MIT Licensed Core.