Cognitive Profiles¶

Cognitive profiles are pre-configured scoring presets that modulate how the memory system prioritizes, retrieves, and consolidates information. They act as a thalamic filter — adjusting the balance between similarity-driven and importance-driven recall to match different task contexts.

How Profiles Work¶

Every recall query is scored using the fused cognitive score formula:

Where:

• α (alpha) — Weight on vector similarity (how close is this memory to the query?)
• β (beta) — Weight on learned importance (how important was this memory at ingestion?)
• α + β = 1.0 — Always normalized

A profile sets α, β, and optional modifiers (hyperfocus boost, lateral mode, episode pinning) to bias the scoring pipeline for a specific cognitive strategy.

Built-in Profiles¶

Standard Profiles¶

Advanced Profiles — Neurodivergent¶

These profiles go beyond α/β tuning — they activate specialized scoring mechanics in the 6-Phase Pipeline and model specific neurocognitive patterns.

Self-Tuning Retrieval (profile=auto)¶

When profile is set to "auto", Spector activates the ProfileAdaptor contextual bandit. It deterministically hashes the current synaptic filter tags and queries its reinforcement stats.

• Epsilon-greedy exploration: The system selects the best performing profile for the tag context 90% of the time, and randomly explores alternative profiles 10% of the time to avoid local optima.
• Cold start fallback: If there are fewer than 10 reinforcement signals for a tag context, it falls back to the configured SalienceProfile default profile, and finally to BALANCED.
• Durable reinforcement: Learned stats are snapshotted to the coactivation.tracker (COAX v2) file on engine shutdown.

New Profile Deep Dives¶

PARANOID_SENTINEL — Amygdala Threat Detection¶

Biological analog: The amygdala's threat-detection circuitry, which filters sensory input for potential dangers and amplifies recall of negative experiences (mood-congruent memory bias).

Use case: SRE agents, security auditors, compliance monitors. Only surfaces memories associated with negative outcomes — errors, failures, security incidents, regressions.

How it works:

• Only negative memories pass the valence filter in Phase 3 of the scorer. Successes, neutral logs, and positive outcomes are invisible.
• Importance-dominated — the severity of the past failure matters more than how closely it matches the current query.
• Query valence is set to -128 (maximum threat), triggering mood-congruent recall amplification.

THE_EXECUTOR — Prefrontal Executive Function¶

Biological analog: The prefrontal cortex in full executive function mode — goal-directed, no tangential exploration, pure task completion.

Use case: Devin-style agentic task runners. Combined with Zeigarnik Effect (markUnresolved()) for tracking open tasks that resist decay.

How it works:

• Heaviside Cliff scoring: The strictness coefficient reshapes the similarity curve into a cliff function:

At strictness=1.0 (default), this is a gentle hyperbola. At strictness=10.0, it's a cliff — 95% of candidates score near zero, and only the closest matches survive.

• Lateral retrieval disabled: No DIVERGENT-style cross-domain exploration. Results must be directly relevant.
• Zeigarnik integration: Unresolved tasks (flagged via markUnresolved()) resist time-decay entirely — their decay bucket is clamped to 0.

HIGHLY_SENSITIVE — Sensory Processing Sensitivity¶

Biological analog: Enhanced sensory processing depth (Aron & Aron, 1997). The highly sensitive brain processes stimuli more deeply, captures finer environmental details, and has a lower threshold for emotional activation.

How it works:

• Lower flashbulb threshold (2.0 vs 3.0): Captures more "important" moments as flashbulb memories. Events that BALANCED would consider routine, HIGHLY_SENSITIVE pins permanently.
• Stronger lateral inhibition (0.3 floor): Less interference between memories. Each memory maintains its distinctiveness rather than blurring with similar neighbors.
• minImportance=0.01: Nothing is too small to remember. Subtle signals that other profiles would round down to zero are preserved.

DEFAULT_MODE_NETWORK — "Shower Thoughts"¶

Biological analog: The brain's default mode network (DMN), which activates during rest, mind-wandering, and unfocused cognition. The DMN surfaces deep, consolidated knowledge rather than recent events.

How it works:

• Skips Working and Episodic tiers entirely. Only Semantic (consolidated facts) and Procedural (learned procedures) are searched.
• α=0.2, β=0.8: Importance-dominated. The DMN isn't looking for direct matches — it surfaces whatever the agent "knows deeply" about a topic.
• No recency bias: Since Episodic is skipped, all results are from long-term consolidated memory. No "what happened today" noise.

EXECUTIVE_DYSFUNCTION — Hebbian-First Associative Recall¶

Biological analog: Prefrontal cortex executive dysfunction. When top-down goal-directed query formulation struggles, bottom-up associative retrieval via the hippocampus and basal ganglia remains intact. Memories surface through directed Spike-Timing-Dependent Plasticity (STDP) causal chains rather than vector query matching.

Use case: Agents struggling with ambiguous inputs or unable to formulate clear semantic queries, relying instead on associative context transitions.

How it works:

• Hippocampal Replay Seed: Rather than searching vectors from a query, the system extracts the last 20 context tags from a sliding RecallHistory buffer.
• STDP Predictive Association: It queries the CoActivationTracker for directed STDP edges leading from those recent context tags, predicting the next logical tag set.
• Tag-Gated Semantic Recall: It runs a standard recall scan using these predicted tags as a strict synaptic Bloom filter.
• STDP Rescoring: Retrieved candidates are boosted based on their Hebbian predictive strength and recency decay.

Usage¶

Via Profile Preset¶

memory.recall("database deadlock", profile: HYPERFOCUS)

Via Recall Options¶

memory.recall("performance optimization",
    profile: DIVERGENT,
    topK: 20,
    lateralDistanceThreshold: 1.5)

Via MCP Tool¶

{
  "name": "memory_recall",
  "arguments": {
    "query": "database deadlock",
    "profile": "HYPERFOCUS",
    "top_k": 10
  }
}

Profile Selection Guide¶

flowchart TD
    A["What is the agent doing?"] --> B{"Focused on\none topic?"}
    B -->|Yes| C{"Need encyclopedic\ndetail?"}
    C -->|Yes| D["SYSTEMATIZER"]
    C -->|No| E["HYPERFOCUS"]
    B -->|No| F{"Exploring new\nterritory?"}
    F -->|Yes| G{"Want cross-domain\ninsights?"}
    G -->|Yes| H["DIVERGENT"]
    G -->|No| I["EXPLORING"]
    F -->|No| J{"Task execution\nor debugging?"}
    J -->|"Executing tasks"| J2["THE_EXECUTOR"]
    J -->|"Debugging"| K["DEBUGGING"]
    J -->|"Threat hunting"| M["PARANOID_SENTINEL"]
    J -->|"Need deep insight"| N["DEFAULT_MODE_NETWORK"]
    J -->|"Detail-sensitive"| O["HIGHLY_SENSITIVE"]
    J -->|No| L["BALANCED"]

Agent Self-Extension¶

Agents can dynamically switch profiles during a conversation:

1. Start with BALANCED for general context
2. Switch to HYPERFOCUS when a specific topic is identified (e.g., user mentions "database deadlock")
3. Switch to DIVERGENT when stuck — lateral results may surface unexpected solutions
4. Switch to SYSTEMATIZER when building a comprehensive knowledge base

The hyperfocus system supports TTL-based activation with agent self-extension:

flowchart LR
    DETECT["Agent detects<br/>focused topic"] --> ACTIVATE["Activate hyperfocus<br/><i>tags: database, deadlock</i>"]
    ACTIVATE --> BOOST["Matching memories<br/>get boost multiplier"]
    BOOST --> CHECK{"Topic continues?"}
    CHECK -->|"Yes"| EXTEND["Extend TTL"]
    CHECK -->|"No — TTL expires"| DEACTIVATE["Auto-deactivate<br/><i>default: 30 min</i>"]

    style ACTIVATE fill:#e74c3c,color:white
    style DEACTIVATE fill:#95a5a6,color:white

Result Metadata¶

Each result carries a retrieval mode indicating how it was retrieved:

Agents can use this metadata to adjust their reasoning — for example, treating LATERAL results with more caution, or presenting HYPERFOCUS results with higher confidence.

What's Next¶

• Focus Mode — Deep dive on HYPERFOCUS and SYSTEMATIZER
• Explorer — Lateral Retrieval — Cross-domain dual-heap mechanics
• Importance Fusion (ICNU) — Sigmoid-gated importance with dopaminergic I×N interaction
• Synapse — Tags & Scoring — Versioned header layouts (V1/V2/V3) and arousal-modulated decay
• Hebbian — Association Learning — STDP with directed causal edges
• Labs — Research Roadmap — Neuromodulatory Gain, Executive Dysfunction Profile