memory-informed-debug

Memory-Informed Debug

Debugging flow with memory baked in as a non-optional reflex.

Instructions

Step 1: Detect entities

Identify the relevant entity/entities (failing test, failing module, error source) per skills/shared/entity-detection.md.

Step 2: Consult memory (required)

Call pensyve_recall:

• query: a short description of the failure, including secondary entity names where known (e.g., "hybrid-router threshold regression phase-4.3" rather than full stack traces)
• entity: primary detected entity
• types: ["procedural", "episodic"] (debugging benefits most from known-good diagnostic procedures and prior incident outcomes)
• limit: 5

Secondary entities are folded into the query string since the MCP server scopes results by single primary entity only.

Surface one line: Recalled N memories from prior debug sessions on <entity>. (Skip if N=0.)

If a highly similar incident is found (score >0.8), call that out inline: This looks similar to an incident captured on <date>: <summary>. Consider that path first.

Step 3: Diagnose

Proceed with the diagnostic work. Use recalled procedural memories as a starting sequence; update the sequence when you learn something new.

Step 4: Capture lesson (when it lands)

When a root cause is confirmed (not just hypothesized), call the memory reflex per skills/shared/memory-reflex.md:

• Episodic — pensyve_observe with episode_id: <session episode_id>, source_entity: "claude-code", about_entity: <primary_entity>, content: "[proactive/in-flight/tier-1] <one-sentence root cause>", content_type: "text".
• Procedural — if the debug produced a reusable diagnostic sequence, capture it: pensyve_observe with episode_id: <session episode_id>, source_entity: "claude-code", about_entity: <primary_entity>, content: "[procedural] [proactive/in-flight/tier-1] trigger=..., action=..., outcome=...", content_type: "text".
• Semantic — if the root cause reveals a durable truth (e.g., "our CI runner's Python 3.10 cannot handle the new syntax"), also call pensyve_remember.

Surface one line: ↳ captured: <one-sentence>.

Step 5: Capture abandoned approach

If an approach was tried and abandoned, observe that too. These are high-value — they prevent re-treading the same dead end next time.

Constraints

• Recall is not optional. If you skip recall to save time, you defeat the working-memory substrate.
• Observe when a lesson lands, not when you first hypothesize.
• One in-flight capture per turn. Additional candidates queue for the Stop residual flush.
• Respect the plugin's noise budget and auto_capture mode.
• Never store secrets or sensitive paths.