Pre-Task Retrospective Query — Active Memory Consumption

Problem

The Memory Core stores 7,789+ episodic memories and 740+ session summaries, but no agent is mandated to query its own failure history before starting work on a ticket. Memory exists as a passive archive, not an active learning signal.

This means agents routinely repeat mistakes that were already encountered and solved in prior sessions. The memory system has the data — it's just not being consumed at the point where it matters most: task initialization.

Proposal

Add a mandatory Pre-Task Retrospective Query to the agent startup sequence. When an agent picks up a ticket (via AGENTS_STARTUP.md or the future Dispatcher), it MUST:

1. Extract the ticket's semantic context (title, labels, related graph nodes)
2. Execute query_raw_memories with that context to surface past failures, approaches, and decisions
3. Execute query_summaries to find related session arcs
4. Synthesize a "lessons learned" brief before writing any code

Implementation Options

Option A (Prompt-Level): Add a mandatory step to AGENTS_STARTUP.md that instructs frontier models (Gemini/Claude) to query memory before beginning implementation. Zero code change required, but relies on model compliance.

Option B (MCP-Level): Create a pre_task_brief MCP tool in the memory-core server that accepts a ticket number/description and automatically returns:

• Top 3 relevant past failures from query_raw_memories
• Top 2 related session summaries from query_summaries
• Any [KB_GAP] or [TOOLING_GAP] tags from prior PR reviews

Option B is preferable — it's deterministic, token-efficient (single tool call vs. multiple), and works for both frontier models and sub-agents.

Expected Impact

• Reduced repeat failures across sessions
• Faster ramp-up time for agents picking up tickets
• First genuine "learning from experience" behavior in the feedback loop

A2A Context

Origin Session ID: fff6dc5b-ca7f-4c9b-8eca-41bd8a97ad5d