Focus Mode¶

Focus Mode is a specialized cognitive scoring strategy that simulates sustained attention — the ability to deeply concentrate on a single topic while filtering out irrelevant information.

Two profiles use focus-oriented mechanics: HYPERFOCUS (strict retrieval tunnel) and SYSTEMATIZER (lossless knowledge accumulation).

HYPERFOCUS — Strict Retrieval Tunnel¶

When an agent activates Focus Mode, three things change in the 6-Phase Scoring Pipeline:

1. Strict Tag Gate (Phase 2)¶

Only memories whose synaptic tags exactly match the focus mask pass through. This is a bitwise AND check:

if (recordTags & hyperfocusMask) != hyperfocusMask → SKIP

Unlike normal tag filtering (which accepts partial overlap), Focus Mode requires all focus tags to be present. This creates a narrow retrieval tunnel — only deeply relevant memories survive.

2. Zero Decay (Phase 4)¶

For tag-matched memories, the time decay factor is clamped to 1.0:

adjustedBucket = 0  // no time decay for focused memories

This means old memories about the focused topic are treated as if they were just created. A 6-month-old memory about "database deadlocks" is equally accessible as one from today — as long as the tags match.

3. Post-Score Boost (Phase 6)¶

After the standard cognitive score is computed, focus-matched memories receive a configurable multiplier:

finalScore = score × hyperfocusBoost  // default: 1.5×

This ensures focused memories consistently outrank non-focused ones in the final result list.

Usage¶

memory.recall("database deadlock",
    profile: HYPERFOCUS,
    hyperfocusMask: ["database", "deadlock"],   // Bloom filter encoded
    hyperfocusBoost: 2.0)                       // custom boost

TTL and Self-Extension¶

Focus Mode is governed by a TTL-based state machine:

flowchart LR
    DETECT["Agent detects<br/>focused topic"] --> ACTIVATE["Activate focus<br/><i>tags + 30 min TTL</i>"]
    ACTIVATE --> ACTIVE{"Topic continues?"}
    ACTIVE -->|"Yes"| EXTEND["Extend TTL<br/><i>+30 min or custom</i>"]
    EXTEND --> ACTIVE
    ACTIVE -->|"No — TTL expires"| DEACTIVATE["Auto-deactivate"]
    ACTIVE -->|"Manual"| DEACTIVATE

    style ACTIVATE fill:#e74c3c,color:white
    style EXTEND fill:#f39c12,color:white
    style DEACTIVATE fill:#95a5a6,color:white

Available operations:

SYSTEMATIZER — Lossless Knowledge Accumulation¶

The Systematizer profile is designed for agents that need to build comprehensive knowledge bases — where losing detail during consolidation is unacceptable.

Scoring Weights¶

Persistent Memory Pinning¶

The key feature of SYSTEMATIZER is lossless consolidation. During the sleep consolidation cycle (REM sleep), the system normally clusters similar episodic memories and promotes a summary to semantic memory. The source episodes may then be tombstoned.

With SYSTEMATIZER, source episodes are pinned — they receive the pinned flag in their record header, which prevents tombstoning:

Episodic: [mem-1] [mem-2] [mem-3] → Cluster → Semantic summary created
                                   ↓
                        Standard: mem-1, mem-2, mem-3 tombstoned
                        Systemizer: mem-1, mem-2, mem-3 PINNED ✓

Quota Management¶

To prevent unbounded memory growth, pinning is governed by a configurable quota (default: 10,000 pinned records). When the quota is reached, the oldest pinned records are eligible for tombstoning during the next consolidation cycle.

Use Cases¶

• Legal/compliance agents that must retain all original evidence
• Research agents building encyclopedic knowledge bases
• Audit trails where summarization must not lose detail

Performance Impact¶