Thermodynamic Physics
Sulcus treats the LLM context window as a bounded physical budget governed by Map vs Territory paging, Dynamic Skills VMMU stubs, interaction-epoch clocks, spaced-repetition stability, and 2-hop graph resonance.
Proactive Paging Lifecycle
Instead of dumping thousands of raw chunk tokens into every prompt, Sulcus separates compact semantic pointers (the Map) from full source artifacts and skill playbooks (the Territory).
PreInvocation Hydration
Before token generation, sulcus hook queries /api/v1/agent/hot-context to inject budgeted <sulcus-memories>, <sulcus-peers>, and skill stubs.
Map Pointer Scan
The model evaluates compact pointer summaries and Dynamic Skills VMMU headers inside its working context budget.
Territory Page Fault
When line-level precision is required, MCP expand_memory or use_skill loads exact sections and reinforces heat by +0.15.
Single Pass Ingest
At turn completion, /api/v1/agent/ingest embeds once, prepares via MemPalace, gates via System 1 VCR, and fires triggers.
Dynamic Skills VMMU
sulcus skill ingest participate in the same Map vs Territory paging discipline through skills_index. Agents see ultra-compact skill capability pointers upfront and call sulcus skill fault (or MCP use_skill) only when executing that specific workflow.Clocks And Stability
If your engineering team takes a long weekend off, project memories should not evaporate by Monday morning. Sulcus defaults tenants to interaction-epoch decay rather than naive wall-clock timers.
Interaction Epoch
decay_mode = "interaction") only when agents store, search, or ingest turns in your tenant. Weekends and holidays cause zero decay.Wall Clock Time
half_life_secs via decay_mode = "time"). Ideal for strictly time-expiring operational telemetry.Hybrid Clock
min(time_decay, interaction_decay) on every tick so high-velocity agent bursts and long calendar gaps both cool stale episodes.effective_half_life = profile.half_life_secs * decay_class.multiplier() * stability
decay_factor = 2^(-tick_interval_secs / effective_half_life)
# Applied on each tick (is_pinned and decay_class="permanent" nodes never decay):
next_heat = max(node.min_heat ?? profile.floor, current_heat * decay_factor)| Archetype | Wall Half Life | Interaction Ops | Heat Floor | Recall Boost | Stability Gain |
|---|---|---|---|---|---|
| episodic | 86,400s (24h) | 50 ops | 0.01 | +0.20 | 1.3x |
| semantic | 2,592,000s (30d) | 500 ops | 0.05 | +0.10 | 1.5x |
| synthesis | 5,184,000s (60d) | 800 ops | 0.05 | +0.10 | 1.5x |
| preference | 7,776,000s (90d) | 1,000 ops | 0.10 | +0.15 | 1.8x |
| procedural | 15,552,000s (180d) | 2,000 ops | 0.08 | +0.10 | 2.0x |
| fact | 31,536,000s (365d) | 5,000 ops | 0.15 | +0.05 | 2.5x |
Node Retention Overrides
decay_class (volatile = 0.5x, normal = 1.0x, persistent = 2.0x, permanent = infinity), a custom min_heat floor (0.0..1.0), and is_pinned = true (exempt from decay and prioritized in active working memory).Edge Resonance Tuning
When a memory is recalled or boosted, thermal energy propagates across connected Knowledge Graph edges up to depth = 2 hops (spread_factor = 0.30, damping = 0.50 per hop, thermal_gate = 0.05).
# Inspect or customize your tenant ThermoConfig
curl -X PATCH https://api.sulcus.ca/api/v1/settings/thermo \
-H "Authorization: Bearer $SULCUS_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"decay_mode": "interaction",
"decay_profiles": {
"episodic": { "half_life_secs": 86400, "half_life_interactions": 50, "floor": 0.01, "reinforce_on_recall": 0.20, "stability_gain": 1.3 },
"semantic": { "half_life_secs": 2592000, "half_life_interactions": 500, "floor": 0.05, "reinforce_on_recall": 0.10, "stability_gain": 1.5 },
"synthesis": { "half_life_secs": 5184000, "half_life_interactions": 800, "floor": 0.05, "reinforce_on_recall": 0.10, "stability_gain": 1.5 },
"preference": { "half_life_secs": 7776000, "half_life_interactions": 1000, "floor": 0.10, "reinforce_on_recall": 0.15, "stability_gain": 1.8 },
"procedural": { "half_life_secs": 15552000, "half_life_interactions": 2000, "floor": 0.08, "reinforce_on_recall": 0.10, "stability_gain": 2.0 },
"fact": { "half_life_secs": 31536000, "half_life_interactions": 5000, "floor": 0.15, "reinforce_on_recall": 0.05, "stability_gain": 2.5 }
},
"resonance": {
"spread_factor": 0.30,
"depth": 2,
"damping": 0.50,
"thermal_gate": 0.05
}
}'