Dashboard Guide
The Overview dashboard at /dashboardgives you a real-time view of your memory graph's health, performance, and activity. Every card tells a story about how your agent's memory is performing.
Sulcus · Dashboard Reference · 2026
Top-Level Stats
The four headline numbers at the top of your dashboard
The total number of memory nodes in your graph. This includes all types — episodic, semantic, procedural, preference, and fact. Pinned and unpinned. A growing node count means your agent is building knowledge. A plateau means consolidation is keeping pace with creation.
The number of API requests processed in the current period. Includes stores, recalls, boosts, and other operations. A healthy graph sees a balance of stores and recalls — all stores and no recalls means the agent is hoarding without using.
The average heat across all memory nodes. A healthy graph typically sits between 0.30–0.60. If average heat is very high (>0.80), memories aren't decaying properly or too many are pinned. If very low (<0.15), the agent isn't recalling enough to reinforce knowledge.
The average response time for memory operations in milliseconds. Includes store, recall, and search latency. Under 100ms is excellent. Over 500ms may indicate the graph is too large for current index settings or the active index needs tuning.
Memory Types Distribution
The breakdown of your graph by memory classification
This chart shows how many memories exist per type. The distribution tells you about your agent's memory patterns — heavy on episodic means lots of event-logging, heavy on procedural means strong operational knowledge.
episodicEvents, conversations, and moments. Fast-decaying by design. These are the "what happened" memories — session logs, interaction records, timestamped observations.
semanticKnowledge and concepts. The "what is" memories — definitions, explanations, domain knowledge that an agent learns from conversations or documents.
proceduralHow-to knowledge and workflows. The "how to do" memories — deployment steps, build processes, troubleshooting guides. Long half-life because operational knowledge should persist.
preferenceUser opinions and settings. The "what they like" memories — preferred tools, communication style, timezone, coding conventions. Higher floor ensures these stay visible.
factVerified data points. The "what is true" memories — API endpoints, version numbers, configuration values, names, dates. Longest half-life and highest floor because facts should be reliable.
Healthy distribution: Most agents should have a mix of all types. If 90%+ of your memories are episodic, your agent is logging events but not distilling knowledge. Consider adding semantic and procedural memories to capture lessons learned from those episodes.
Heat Distribution
How your memories are distributed across temperature bands
The heat distribution chart groups all memories into five temperature bands. This gives you an at-a-glance view of graph vitality — are most memories hot and active, or cold and fading?
| Band | Heat Range | Color | Meaning |
|---|---|---|---|
| Blazing | 0.80 – 1.00 | 🔴 Red | Recently created or heavily recalled. These memories are at the top of the active index and guaranteed to be in context. |
| Hot | 0.50 – 0.79 | 🟠 Orange | Active and healthy. Being recalled periodically or recently created. Well within the active index. |
| Warm | 0.25 – 0.49 | 🟡 Yellow | Cooling but still visible. May appear in the active index depending on max nodes setting. Should be recalled soon if important. |
| Cool | 0.10 – 0.24 | 🔵 Blue | Approaching the cold threshold. Unlikely to be in the active index unless max nodes is very high. At risk of consolidation. |
| Frozen | 0.00 – 0.09 | ⚪ Gray | Below or near the cold threshold. Consolidation candidates. Only the floor (min_heat) keeps them from reaching absolute zero. |
What to look for:A healthy graph has a natural bell curve — some blazing (recent), most warm/hot (active knowledge), some cool/frozen (old/unused). If everything is frozen, the agent isn't recalling. If everything is blazing, decay may be too slow or too many memories are pinned.
Summary Cards
Quick-reference cards for each engine subsystem
Below the main charts, four summary cards show the current configuration and status of each thermodynamic subsystem. These mirror the settings at /dashboard/account — clickable shortcuts to the full configuration panel.
Shows whether the decay tick is active or paused. When active, the engine runs decay calculations on a regular interval. Pausing the tick freezes all heat values — useful during migrations or debugging.
Summarizes current resonance settings: spread factor, damping, depth, and thermal gate. Shows whether resonance is enabled and the effective propagation radius of each recall event.
Shows max nodes and context budget, plus the current utilization — how many nodes are actually being injected and what percentage of the character budget is used. If utilization is consistently at 100%, consider increasing the limits.
Shows the aggregate recall accuracy over the last 30 days. Breaks down by memory type so you can identify which types are surfacing well and which need tuning. See the Thermodynamic Engine guide for tuning advice.
Graph Health
The vital signs of your memory graph
The Graph Health panel provides a deeper diagnostic view of your memory graph's condition. These metrics help you identify imbalances and optimize performance.
Count of permanently pinned memories that are exempt from decay. High pin counts mean more memory is always in context — good for critical knowledge, bad for context budget if overdone.
The mean heat across all nodes — same as the top-level stat but shown here with historical trend. A declining average suggests the agent isn't recalling enough.
The single hottest node in the graph, with its label and heat score. Useful for spotting runaway heat — a memory that's being recalled too frequently or has excessive stability.
The coldest non-floor node. If this memory is important, it needs recall or pinning before consolidation claims it.
The number of distinct agent namespaces writing to this graph. Multi-agent setups show all contributors — useful for understanding which agents are most active.
Recent Activity
A live feed of the latest memory operations
The Recent Activity feed shows the last memory operations in chronological order. Each entry shows the operation type (store, recall, boost, relate, decay tick), the memory label, namespace, heat change, and timestamp.
Use this feed to understand your agent's behavior in real time. If you see rapid-fire stores without any recalls, the agent is writing but not reading. If you see frequent boosts on the same memory, it's being heavily relied upon and the stability multiplier is climbing.
Example activity feed — your dashboard shows live data
Your graph at a glance.
The dashboard is your window into how memory is performing. Check it periodically to ensure your agents are building and recalling knowledge effectively.