Memory Detectors

Memory detectors are algorithms that analyze your command history to surface patterns. They run on boot and after each command execution.

Detector A: Preferred Working Directory

Function: detectPreferredCwd

Identifies directories where you spend significant time.

Trigger: 5+ command executions across 2+ sessions in the same directory.

Output: a suggestion with kind preferred_cwd, key = the directory path, confidence 0.70–0.95.

Confidence scaling: base 0.70, increases with execution count. More executions in more sessions = higher confidence, capped at 0.95.

Why it matters: the planner uses preferred CWDs to set context. If you always work in ~/projects/myapp, the planner knows to interpret “run tests” in that context.

Detector B: Recurring Commands

Function: detectRecurringCommands

Identifies commands you run frequently.

Trigger: 4+ executions of the same command family. Command families are normalized — npm test, npm run test, and npm t may be grouped if they share a prefix.

Output: a suggestion with kind recurring_command, key = the command, confidence 0.60–0.90.

Confidence scaling: base 0.60, increases with frequency across sessions. More varied sessions = higher confidence.

Why it matters: the planner can suggest familiar tools. If you always use rg instead of grep, the planner learns to generate rg commands.

Detector C: Workflow Patterns

Function: detectWorkflowPatterns

Identifies command sequences you repeat — pairs and triples of commands that always appear together in order.

Trigger: 3+ occurrences of the same sequence across 2+ sessions.

Output: a suggestion with kind workflow_pattern, key = the sequence label, confidence 0.65–0.85.

Triple priority: when a three-step sequence (A → B → C) is detected, it suppresses the constituent pairs (A → B, B → C) to avoid redundant suggestions.

Why it matters: these suggestions can be promoted directly to multi-step workflows, automating your most common sequences.

Detector coordination

All three detectors run through runDetectors(), which:

1. Filters history to successful executions only (failed commands are noise)
2. Runs each detector independently
3. Deduplicates against existing memory items and suggestions
4. Returns new suggestions only

Suggestion lifecycle

History data → Detector runs → Suggestion created (pending)
                                        ↓
                              User sees suggestion card
                                        ↓
                    Accept → MemoryItem created → feeds planner
                    Dismiss → suggestion removed permanently

Testing

Detectors are pure functions with comprehensive unit tests in packages/domain/src/memoryDetectors.test.ts. Tests verify threshold behavior, confidence scaling, cross-session requirements, deduplication, and triple-over-pair priority.