What you should see
- Organ cards — one card per domain. More intense color = more recent activity. Click to explore.
- Heart Rate (total store + retrieve operations this session) — shown as a red number. ♥ 12 ops means 12 operations happened.
- Memory Pressure (how full the memory budget is, as a percentage) — green below 60%, yellow 60-80%, red above 80%.
- Assoc. Temp (average strength of links between memories) — higher means memories are more interconnected.
Example 1 — Fresh Start
You see one organ card labeled "default" with 3 memories and 0 edges. The heart rate shows ♥ 3 ops. Pressure is at 0.1%.
What this means: You've just started using Fractal Memory. Only 3 memories have been stored, none are linked to each other yet (0 edges means the system hasn't seen any memories co-retrieved together). This is completely normal for a new domain — it's in "PIONEER" stage (fewer than 10 memories).
Example 2 — Healthy Active System
You see 3 organ cards: "work-project" (142 memories, 89 edges), "personal" (45 memories, 23 edges), "default" (8 memories). Heart rate is pulsing at 24 ops. Pressure is 34%.
What this means: The organism has grown 3 distinct domains. "work-project" is the most mature with many memories and a dense association network (89 edges means memories are well-connected — the system has learned relationships between them). 34% pressure means plenty of room for new memories. The high heart rate indicates an active session.
Example 3 — Pressure Warning
One domain "research" shows 487 memories with pressure at 87%. The pressure indicator is red.
What this means: This domain is approaching its memory budget limit (the maximum number of memories it can hold before it starts automatically deleting weak ones). Above 80%, the system prunes aggressively — old, unimportant memories will be deleted via apoptosis to make room. You might want to check the Apoptosis page to see what's being removed, or increase the budget in Config.
What you should see
- Node size — larger circles = more frequently retrieved memories (they're more important).
- Node color: ● Orange = Probation (new, unverified), ● Blue = Confirmed (retrieved enough to be trusted), ● Purple = Permanent (core knowledge, never auto-deleted), ● Red = Safety Critical (must never be lost).
- Edge thickness — thicker lines = stronger association. Edges form from co-retrieval (when two memories are fetched in the same query), spreading activation (when activating one memory activates its neighbors), and REM recombination (when the system "dreams" and combines related memories).
- Clusters — groups of tightly connected nodes represent topic sub-groups within the domain.
Example 1 — Sparse New Domain
5 small orange nodes scattered with no edges between them.
What this means: All 5 memories are new (orange = probation) and have never been retrieved together, so no associations exist yet. This is normal for a brand-new domain. As you start querying and the system retrieves combinations of these memories, edges will form naturally.
Example 2 — Two Clear Clusters
Two tight groups of blue nodes connected internally with thick edges, with only 1-2 thin edges between the groups.
What this means: Your domain has two distinct sub-topics. The thick internal edges show strong associations within each cluster (memories within the same topic are frequently retrieved together). The thin edges between clusters are weak cross-topic links. This is healthy — the Niche Construction page might suggest splitting this into two separate domains.
Example 3 — One Central Hub
One large purple node in the center connected to 15+ smaller blue nodes radiating outward.
What this means: The large purple node is a permanent, frequently-retrieved memory that serves as a "hub" connecting many other concepts. This is typical for core facts (like "this project uses TypeScript" in a dev domain) that get pulled into many different queries. The hub pattern is healthy — it means the system has identified a central piece of knowledge.
What you should see
- ● Orange ring — Probation memories. New and unverified. Must be retrieved within
pruning_ttl_sessions (number of sessions before an unused memory is deleted, default ~5) or they expire.
- ● Blue ring — Confirmed memories. Retrieved at least
confirmation_threshold (default 2) times. Stable and protected from pruning.
- ● Purple ring — Permanent memories. Retrieved
permanence_threshold (default 10) times, or manually pinned. Always loaded at session start.
- ● Red ring — Safety Critical. Manually marked. Always present in every session regardless of topic.
Example 1 — New Domain (All Orange)
The ring is almost entirely orange with tiny blue sliver. Promotion rate: 8%.
What this means: Nearly all memories are still in Probation — this is a young domain where memories haven't been retrieved enough to confirm. The 8% promotion rate is low but expected. As you interact more with this domain, memories will naturally promote as they prove useful.
Example 2 — Mature Healthy Domain
Ring shows 20% orange, 55% blue, 20% purple, 5% red. Promotion rate: 72%.
What this means: Excellent health. Most memories (55%) are Confirmed — trusted and stable. 20% Permanent means the system has identified core knowledge. The 20% orange is a healthy churn of new information being evaluated. 72% promotion rate means the gate is selecting well — most stored memories turn out to be useful.
Example 3 — Stagnant Domain
Ring shows 85% orange, 12% blue, 3% purple. Promotion rate: 15%.
What this means: Too many memories stuck in Probation. This usually means: (a) memories are being stored but never queried, so they can't confirm, or (b) the gate is too loose, storing low-quality information. Check the Feedback page for retrieval patterns, and consider tightening the gate via the Processing Pipeline page.
What you should see
- Waiting Room — list of candidate texts with their detected signals. Entries expire after
waiting_room_expiry_sessions sessions.
- Gate Signal Count — the threshold number. 1 = Pioneer (accepts easily), 2 = Standard, 3+ = Strict.
- Signal types — emphasis (marked important), explicit_store (user asked to remember), repetition (mentioned again), novelty (new information), topic_relevance (matches current context).
- Throughput — memories stored per session. Healthy range depends on domain activity.
Example 1 — Bootstrapping
Gate signal count is 1. Waiting room has 0 entries. 8 memories stored this session.
What this means: Pioneer mode is active — every piece of information that gets even one signal passes through immediately. The empty waiting room confirms nothing is stuck. 8 memories stored is aggressive but appropriate for a new domain that needs to build its initial knowledge base quickly.
Example 2 — Healthy Production
Gate signal count is 2. Waiting room has 3 entries (ages: 1 session, 2 sessions, 4 sessions). 2 memories stored this session.
What this means: Standard selectivity. 3 candidates are waiting for a second confirming signal. The 4-session-old entry will expire soon if it doesn't get reinforced — meaning it probably wasn't important enough. 2 stores per session is a healthy, selective rate.
Example 3 — Gate Too Strict
Gate signal count is 3. Waiting room has 28 entries. 0 memories stored in the last 5 sessions.
What this means: The gate requires 3 independent signals but most information only gets 1-2. Memories are piling up in the waiting room and expiring without ever being stored. Lower the gate_signal_count to 2 (or even 1 temporarily) in Config to allow learning.
What you should see
- System State — FLOW (green, balanced), STRESS (red, defensive mode), CURIOSITY (cyan, exploring).
- Trajectory chart — 3 colored lines showing signal levels over sessions. Stable flat lines = mature domain. Spikes = unusual activity.
- Response curves — 8 S-shaped curves. Each maps a signal to a parameter. The midpoint (where the curve hits 50%, the "tipping point") and steepness (how sharply it transitions) are learned.
- Convergence gauges — green = Converged (variance < 0.01), yellow = Learning, red = Unstable.
Example 1 — Early Learning
State shows CURIOSITY (cyan). All 8 convergence gauges show "Learning" in yellow. Trajectory chart shows volatile lines.
What this means: The system is in its initial exploration phase. CURIOSITY state means the domain is sparse and the organism is actively experimenting with different parameter settings. The curves haven't converged because they're still learning the optimal shape. This is expected for domains with fewer than 30 sessions.
Example 2 — Fully Converged
State shows FLOW (green). All 8 gauges show "Converged" in green. Trajectory lines are flat.
What this means: The organism has self-calibrated. All response curves have found stable shapes, meaning the system knows exactly how to translate activity signals into parameter adjustments. FLOW state means everything is balanced — no unusual stress or sparsity. This is the ideal state for a mature domain.
Example 3 — Stress Response
State shows STRESS (red). Retrieval pressure line has spiked. 2 curves show "Unstable". A convergence gauge just flipped from green to yellow.
What this means: Something changed — maybe the user is querying heavily with many misses, or making lots of corrections. STRESS state tightens the storage gate (to avoid storing more bad memories) and boosts retrieval aggressiveness (to try harder to find good matches). The curves de-converged because conditions shifted significantly. Check the Feedback page for what triggered the stress.
What you should see
- Hit Rate gauge — green ≥60% (healthy), yellow 40-60% (needs attention), red <40% (struggling).
- Maturity badge — PIONEER (<10 memories), SHRUB (<50), FOREST (<200), MATURE (200+).
- Parameter Drift — list of parameters being pushed >2× their default value. Red = investigate.
- Last Retrieval Explained — step-by-step breakdown of the most recent query: which memories scored highest, what boosts/penalties were applied, which bandit arm was used.
Example 1 — Pioneer Stage
Maturity: PIONEER. Hit rate: N/A. No drift warnings. "Not enough data yet" for cascade predictions.
What this means: Too early to measure health — the domain has fewer than 10 memories. Most subsystems (morphogens, bandits, niche construction) are not yet active. Focus on storing foundational memories first; health metrics become meaningful after ~20+ memories.
Example 2 — Healthy Mature System
Maturity: FOREST. Hit rate: 73% (green). 0 drift warnings. 0 cascade warnings. Last retrieval shows 3 memories retrieved with high scores.
What this means: The system is performing well. 73% of retrievals returned useful memories. No parameters are drifting abnormally and no subsystem failures are predicted. The last retrieval was clean — the right memories were found with high confidence scores.
Example 3 — Retrieval Struggling
Maturity: FOREST. Hit rate: 28% (red). 2 drift warnings: "gate_threshold" and "retrieval_top_l1". Cascade warning: "retrieval subsystem degraded, severity 0.7".
What this means: Only 28% of retrievals are useful — the system is returning wrong or irrelevant memories most of the time. The morphogen curves are over-compensating (pushing gate_threshold and retrieval parameters far from default). The cascade detector predicts potential retrieval failure. Check Feedback for correction patterns, and consider adding more precise memories to the domain.
What you should see
- FLOW (green) — balanced state. Multipliers near 1.0. System is performing normally.
- STRESS (red) — defensive. Gate tightens (stores less), retrieval becomes more aggressive (tries harder), scar sensitivity increases.
- CURIOSITY (cyan) — exploratory. Gate loosens (stores more), retrieval is less strict, exploration weight increases. Fires automatically in sparse domains.
- Transition History — colored segments per session. Green/Red/Cyan bars showing which state was active.
Example 1 — Stable FLOW
State: FLOW. Miss counter: 0. Success counter: 7. All multipliers between 0.95 and 1.05. History shows solid green for the last 10 sessions.
What this means: Everything is working smoothly. 7 consecutive successful retrievals, no misses. All parameters are at their natural (unmultiplied) values. 10 sessions of continuous FLOW means the domain is stable and well-calibrated.
Example 2 — Triggered STRESS
State: STRESS. Miss counter: 4. Gate multiplier: 0.6. Retrieval aggression: 1.8. History shows: green, green, red, red.
What this means: 4 consecutive misses triggered STRESS. The gate multiplier of 0.6 means the gate threshold is reduced by 40% (harder to store new memories — the system is being cautious about adding more potentially bad information). Retrieval aggression at 1.8 means the system is trying 80% harder to find matching memories. This will resolve once retrievals start succeeding again.
Example 3 — CURIOSITY in New Domain
State: CURIOSITY. Domain maturity: PIONEER. Gate multiplier: 1.5. Exploration weight: 2.0. History shows all cyan.
What this means: The domain is new and sparse, so the system automatically entered CURIOSITY mode. Gate multiplier 1.5 means it's 50% easier to store memories (lowering the bar to build the initial knowledge base). Exploration weight 2.0 means the system is actively trying different retrieval strategies. This is healthy and expected — it will transition to FLOW as the domain matures.
What you should see
- 5 Arms: standard, aggressive, conservative, explorer, precision — each configures retrieval differently (how many results, scoring weights, etc.).
- Win rate bars: green ≥60%, yellow ≥40%, red <40%.
- Posterior (Alpha/Beta) — higher Alpha (successes) relative to Beta (failures) = more confidence that the arm is good.
- Context bins — 27 cells showing which morphogen combinations have been seen. Rare bins show uniform priors (equal probability — system hasn't learned yet).
Example 1 — Uniform Priors (No Data)
All 5 arms show ~50% success rate. Context bins are mostly empty. Alpha/Beta ratios are all near 1:1.
What this means: The system hasn't collected enough data to differentiate between strategies. All arms look equally promising (or equally uncertain). This is the starting state — after 10-20 sessions of active retrieval, the arms will begin to separate.
Example 2 — Clear Winner
"Precision" arm shows 78% win rate (green). "Aggressive" shows 42% (yellow). "Explorer" shows 35% (red). Current arm: Precision in context bin [medium, high, medium].
What this means: The system has learned that the "precision" strategy (returning fewer, higher-confidence results) works best when information richness is medium, retrieval pressure is high, and domain density is medium. It now preferentially selects this arm in similar contexts. The aggressive and explorer arms underperform in this situation.
Example 3 — Post-Dream Reset
All arms shifted closer to 50%. A "Downscaled" badge appears. Previously dominant "conservative" arm dropped from 82% to 65%.
What this means: Dreaming's slow-wave phase just downscaled the posteriors (reduced the influence of historical data to prevent over-commitment to one strategy). This is intentional — it prevents the system from being permanently locked into a strategy that was good in the past but might not be optimal now. The arms will re-differentiate as new data comes in.
What you should see
- Slow Wave stats — edges_decayed (count), edges_pruned (fully removed), bandit_downscale (yes/no).
- REM efficiency donut — green = stored, red = discarded, grey = not attempted.
- Dream History chart — per-session bars showing edges decayed and syntheses stored. Regular cycles with gradual synthesis accumulation = healthy.
Example 1 — First Dream Cycle
Slow wave: 0 edges decayed (none exist yet). REM: 0 attempted (too few memories to pair). Dream history is empty.
What this means: The domain is too new for meaningful dreaming. No associations exist to decay, and not enough memory pairs meet the resonance threshold for REM synthesis. Dreaming becomes productive after ~15-20 memories with some association edges.
Example 2 — Productive Night
Slow wave: 12 edges decayed, 3 pruned. REM: 5 attempted, 3 stored, 2 discarded. Donut shows 60% green.
What this means: Good dream cycle. 12 weak edges were weakened further, and 3 were so weak they were removed entirely. REM found 5 promising memory pairs and successfully synthesized 3 new insight memories (2 didn't meet quality thresholds). 60% REM efficiency is healthy — the system is generating useful new knowledge from existing memories.
Example 3 — Inefficient REM
REM: 12 attempted, 1 stored, 11 discarded. Donut is mostly red (8% efficiency).
What this means: REM is finding many candidate pairs but almost all syntheses are being rejected. The rem_emergence_threshold (quality bar for a synthesized memory to be accepted) is probably too strict. Lower it in Config to allow more creative combinations through, or the domain's memories may be too dissimilar to combine meaningfully.
What you should see
- Stat tiles — edges_strengthened, merges, edges_decayed, scars_expired, freeloaders (memories boosted by association but never directly retrieved — they're "free-riding").
- Merge Events — pairs of memories with cosine similarity (how similar they are, 0-1) above 0.80 that were combined.
- History chart — per-session metrics. Healthy: stable edge strengthening, occasional merges, gradual decay.
Example 1 — Routine Consolidation
Edges strengthened: 8. Merges: 0. Edges decayed: 3. Scars expired: 1. Freeloaders: 2.
What this means: Normal, healthy consolidation. 8 co-retrieval bonds were reinforced, 3 old weak edges were weakened further, 1 expired scar was cleaned up. 2 freeloaders detected — these memories are being surfaced via association links but never directly matched by queries.
Example 2 — Heavy Merging
Merges: 7. The merge events table shows pairs with 0.85-0.92 cosine similarity.
What this means: Too many near-duplicate memories were stored. 7 pairs were close enough to merge. This usually means the gate is accepting variations of the same information (e.g., "use async/await" and "always use async/await for I/O"). Consider raising gate selectivity or merge threshold.
Example 3 — Domain Drift Detected
Domain drift shows 3 memories in "work-project" that are closer to "personal" domain centroid.
What this means: 3 memories were stored in "work-project" but their content is actually more relevant to "personal". Maybe you discussed personal preferences during a work session. The suggestion is advisory — you can manually move them via the API, or let Niche Construction handle it automatically.
What you should see
- Trigger types — context_drift (drifted from domain), contradicted (too many corrections/scars), obsolescence (never retrieved, low vitality).
- Vitality histogram — should cluster near 0. If high-vitality memories are being pruned, something is wrong.
- Recent Events table — pruned memory, reason, vitality at death, fragment count, scar created?
Example 1 — Healthy Pruning
3 memories pruned via obsolescence. Vitality scores: 0.02, 0.05, 0.01. Fragments: 1, 0, 2.
What this means: 3 memories with near-zero vitality (almost never retrieved, very old) were cleaned up. 3 fragments were salvaged — small atomic facts extracted before the full memories were deleted. This is completely healthy — the system is removing dead weight while preserving useful nuggets.
Example 2 — Topic Shift
8 memories pruned via context_drift in one session. All from "web design" sub-topic in a "fullstack" domain.
What this means: The domain's focus has shifted away from web design toward backend topics. Memories about CSS tricks, responsive layouts, etc. have drifted far from the current context centroid. If you still need these memories, consider creating a separate "web-design" domain.
Example 3 — Contradiction Cleanup
2 memories pruned via "contradicted". Scar count on both: 4+. Scars created: yes.
What this means: These memories were corrected 4+ times — they kept returning wrong information. The system finally removed them and created scar tissue (a pattern that suppresses similar future retrievals) so similar bad memories won't surface again. Check the Kintsugi page to see the new scars.
What you should see
- Domain Health Cards — internal cohesion (how strongly memories within a domain link to each other, 0-1) and external separation (how distinct a domain is from other domains, 0-1). Both should approach 1.0.
- Scatter plot — X=cohesion, Y=separation. Top-right = ideal. Bottom-left = messy.
- Proposals — suggested actions with modularity delta (how much the overall organization score improves if applied) and confidence.
Example 1 — Well-Organized
All domains in top-right quadrant. Cohesion: 0.82-0.91. Separation: 0.78-0.88. No proposals.
What this means: Domains are well-separated and internally coherent. Memories within each domain are strongly related to each other but distinct from other domains. No reorganization needed.
Example 2 — Split Proposal
"fullstack" domain at cohesion 0.42, separation 0.65. Proposal: split into "frontend" and "backend" with modularity delta +0.18, confidence 0.84.
What this means: The "fullstack" domain contains two distinct clusters that would work better as separate domains. The low cohesion (0.42) confirms that memories within the domain aren't strongly linked — there are two sub-groups talking past each other. Splitting would improve overall organization by 0.18 points.
Example 3 — Merge Proposal
"css-tricks" and "styling" domains both at low separation (0.31 and 0.28). Proposal: merge with modularity delta +0.12, confidence 0.76.
What this means: These two domains overlap heavily — memories in one are very similar to memories in the other (low separation). Merging them would consolidate related knowledge and strengthen association edges that currently span two domains.
What you should see
- EKF State Gauges — 5 bars for gate, retrieval, storage, associations, domains. Green/Yellow/Red.
- Badge — EKF Active (cyan) = full prediction. Circuit Breaker (yellow) = simplified fallback.
- Covariance Trace — measures uncertainty. Lower = more confident predictions.
- Warnings — affected system, severity (0-1), sessions until predicted impact, recommended action.
Example 1 — All Clear
All 5 gauges green (0.82-0.95). EKF Active badge. Covariance trace: 0.03. 0 warnings.
What this means: All subsystems are healthy and the EKF is confident in its estimates (low covariance). No cascading failures predicted. This is the ideal state.
Example 2 — Retrieval Degradation Warning
Retrieval gauge: yellow (0.52). Warning: "retrieval subsystem degraded, severity 0.6, recommended: increase retrieval_top_l1, impact in 3 sessions".
What this means: The EKF detects that retrieval quality is declining and predicts it will become critical in 3 sessions. The recommended action is to increase retrieval_top_l1 (the number of candidate memories pulled in the first retrieval pass), which widens the search net to find better matches. Act now to prevent a cascade into STRESS state.
Example 3 — Circuit Breaker Fallback
Badge shows "Circuit Breaker" (yellow). No EKF gauges visible. Simple counter: "3 consecutive low hit rate sessions".
What this means: The EKF couldn't be initialized (not enough session history, or dependencies missing). The system fell back to a simpler circuit breaker (just counts consecutive bad sessions and triggers an alert at a threshold). 3 consecutive low hit rate sessions have been detected — this is equivalent to a STRESS trigger. The system will auto-adapt, but EKF would provide more precise guidance.
What you should see
- Regime — BOOTSTRAP (≤5 sessions or <30 memories), LEARNING (morphogens not converged), MATURE (converged + sufficient data).
- Subsystem badges — green = active/running, grey = folded/standby.
- Anomaly History — rolling 10-session window. Spikes trigger forced unfolds.
- Forced Unfold Events — timestamp, which subsystem, what metric triggered it.
Example 1 — Bootstrap
Regime: BOOTSTRAP. Only 3 subsystems active (green): gate, retrieval, storage. 8 subsystems folded (grey).
What this means: The domain is very new. Only the essential subsystems (store, retrieve, basic gate) are active. Advanced features like morphogens, bandits, niche construction, and dreaming are folded — they'd produce meaningless results with so little data. They'll activate automatically as the domain grows.
Example 2 — Mature with All Active
Regime: MATURE. All 11 subsystems active (green). 0 forced unfolds in last 10 sessions.
What this means: Full capability unlocked. The domain has enough data and converged morphogens to run every subsystem. No anomalies detected recently. This is the target state for a well-established domain.
Example 3 — Forced Unfold
Regime: MATURE. "Apoptosis" was folded but just force-unfolded. Anomaly: vitality scores dropped 2.3 standard deviations below mean.
What this means: The apoptosis subsystem was on standby (the domain was stable, no pruning needed). But a sudden drop in vitality scores across many memories triggered an anomaly — something changed, and the system woke up apoptosis to investigate and potentially prune affected memories. Check what caused the vitality drop (maybe a large topic shift).
What you should see
- Signal Stream — live feed. Red badges = DANGER (CORRECTION, RE_ASK, REJECTION). Green badges = SAFETY (CONFIRMATION, ENGAGEMENT).
- Signal Distribution donut — healthy sessions have more green than red.
- Store Triggers — SURPRISE (novel info), RESOLUTION (answer to pending question), PREFERENCE, COMMITMENT, PATTERN.
- Information Richness — 0 to 1 gauge. >0.6 = dense conversation, 0.3-0.6 = moderate, <0.3 = sparse.
Example 1 — Healthy Session
Signal stream: 8 green (CONFIRMATION, ENGAGEMENT), 1 yellow (RE_ASK). Distribution donut: 89% safety. Info richness: 0.72.
What this means: Great session. The system retrieved useful memories 8 times and the user confirmed or engaged positively. Only 1 re-ask means one query wasn't fully answered on the first try. High information richness means the conversation was dense with new facts — the gate should be receptive to storing new memories.
Example 2 — Struggling Session
Signal stream: 4 red (2 CORRECTION, 2 RE_ASK), 2 green. Distribution: 33% safety. Info richness: 0.45.
What this means: The system is struggling. 2 corrections mean wrong memories were surfaced. 2 re-asks mean the user had to repeat themselves. The system will likely transition to STRESS state (see States page). The morphogens will adjust parameters to be more cautious. Consider storing more accurate, specific memories in this domain.
Example 3 — Learning Explosion
Store triggers show 6 SURPRISE, 3 RESOLUTION, 2 PATTERN in one session. Info richness: 0.91.
What this means: Extremely information-dense session. The system detected 6 pieces of novel information (SURPRISE), resolved 3 pending questions, and noticed 2 recurring patterns. With 0.91 info richness, the morphogens will lower the gate threshold to capture more of this valuable content. Expect the lifecycle page to show many new Probation entries after this session.
What you should see
- Buffer composition bars — purple = PINNED, blue = AUTO-PROMOTED, green = JUST WRITTEN, grey = RECENT.
- Access heatmap — bar chart of top 20 memories by access count. "Hot" memories (red/orange) are auto-promotion candidates.
- Habituation indicators — memories with consecutive-turn retrieval get a counter. At the threshold, they're suppressed.
Example 1 — Minimal Buffer
Buffer: 0 pinned, 0 auto-promoted, 2 just written, 3 recent. Total: 5 entries.
What this means: Small working buffer — the session just started. No memories have been pinned or used enough to auto-promote. The 2 "just written" are from the current session's stores, and the 3 "recent" are from the last few retrievals. As the session progresses, the buffer will grow.
Example 2 — Well-Managed Buffer
Buffer: 3 pinned, 5 auto-promoted, 1 just written, 8 recent. Access heatmap shows one memory with 14 accesses.
What this means: 3 manually pinned memories are always in context (likely core domain facts). 5 memories have been accessed often enough to auto-promote. The memory with 14 accesses is clearly important — it's been used heavily. This is a healthy, well-structured buffer.
Example 3 — Habituation Active
One memory shows "habituated (4/4 consecutive turns)". It's crossed out in the buffer list. A different memory now appears in its place.
What this means: A memory was retrieved 4 turns in a row, hitting the habituation window. It's temporarily suppressed to force the system to surface other relevant memories. This prevents a single "sticky" memory from dominating every response. The suppression is temporary — it resets if the memory isn't queried for a few turns.
What you should see
- Tier cards — 5 cards with status: ON TIME, DUE, OVERDUE. Each shows last run timestamp.
- Last Report — action list with counts (e.g., "edges_strengthened: 12, merges: 2, scars_expired: 1"). Duration in ms.
- Actions History — stacked bar chart per run. Spikes = heavy maintenance was needed.
Example 1 — Healthy Schedule
All 5 tiers show "ON TIME" (green). SESSION-END ran 2 hours ago. DAILY ran this morning. Last report: 8 actions, 120ms.
What this means: Everything is on schedule. Maintenance is running at expected intervals, completing quickly (120ms is fast), and performing routine cleanup. No intervention needed.
Example 2 — OVERDUE Daily
DAILY shows "OVERDUE" (red). Last ran 3 days ago. SESSION-END still shows "ON TIME".
What this means: Daily maintenance hasn't run in 3 days. This often happens if the MCP server was restarted frequently (resets the schedule) or if no sessions ran for a few days. SESSION-END still working means dreaming and consolidation are healthy. The daily tasks (deeper cleanup, weekly summaries) are delayed but not critical.
Example 3 — Heavy Maintenance
Last WEEKLY report: 47 actions, 3.2s. Actions: 18 edges pruned, 12 scars expired, 8 apoptosis events, 5 niche proposals, 4 merges.
What this means: Heavy weekly cleanup. 18 weak edges were removed, 12 old scars expired, 8 memories were pruned, and the niche system generated 5 domain reorganization proposals. 3.2s is longer than usual — this domain accumulated a lot of debt. Check Apoptosis and Niche pages for details on what was cleaned up.
What you should see
- Parameter table — category, name, current value, sensitivity badge (H = high, M = moderate, L = low).
- Feature flags — ON / OFF toggles for enable_spreading_activation, enable_scar_penalty, enable_llm_judge, etc.
- Threshold gauges — 8 horizontal bars showing critical thresholds (gate_threshold, confirmation_threshold, etc.) and their current values relative to defaults.
Example 1 — Default Configuration
All parameters at default values. All thresholds centered. 3 feature flags ON (spreading_activation, scar_penalty, llm_judge).
What this means: Fresh system with no morphogen adjustments yet. All parameters are at their baseline values. Core features (association spreading, scar penalties, LLM-based quality judging) are enabled. As sessions accumulate, morphogens will begin shifting H-sensitivity parameters away from defaults.
Example 2 — Morphogen-Tuned
gate_threshold (H) shows current value 0.35 vs default 0.50. retrieval_top_l1 (H) shows 15 vs default 10. Both marked with drift indicator.
What this means: Morphogens have actively tuned two high-sensitivity parameters. The gate threshold was lowered from 0.50 to 0.35 (making it easier to store memories — probably a CURIOSITY or early-stage response). retrieval_top_l1 was increased from 10 to 15 (casting a wider net during retrieval — probably responding to low hit rate). These adjustments are automatic and data-driven.
Example 3 — Troubleshooting
You search "merge" and find merge_similarity_threshold = 0.80. You notice the Consolidation page shows too many merges.
What this means: The merge threshold of 0.80 means any two memories with 80%+ cosine similarity get merged. If too many merges are happening, raise this to 0.85 or 0.90 to be more selective about what counts as a "duplicate". This is a manual configuration change you'd make by modifying FractalConfig.
{% endblock %}