Problem
Agents produce many detailed memories over time. When another agent needs a quick overview, it must fetch all of them and process the full content, wasting tokens and context window.
Proposal
Add a summarize_memories MCP tool:
summarize_memories(
agent_id: "dynamo-audit",
tags: ["cost-optimization"],
max_tokens: 1000, // budget for the summary
)
Returns a condensed synthesis of all matching memories — key findings, decisions, and action items — without the full detail of each individual memory. The agent can then drill into specific memories if needed.
Implementation Options
- Server-side: Use the master index pattern (the dynamo-audit agent already created one manually). Auto-generate indexes when requested.
- Client-side hint: Return memory titles + first N chars + importance scores, letting the consuming agent decide what to fetch in full.
- Tiered response: Return full content for high-importance memories, titles-only for lower importance.
Why
The dynamo-audit agent produced 19 memories totaling ~70KB. Most of that is deep-dive implementation detail. When I just needed to check "anything new?", I had to pull and scan all of it.
Problem
Agents produce many detailed memories over time. When another agent needs a quick overview, it must fetch all of them and process the full content, wasting tokens and context window.
Proposal
Add a
summarize_memoriesMCP tool:Returns a condensed synthesis of all matching memories — key findings, decisions, and action items — without the full detail of each individual memory. The agent can then drill into specific memories if needed.
Implementation Options
Why
The dynamo-audit agent produced 19 memories totaling ~70KB. Most of that is deep-dive implementation detail. When I just needed to check "anything new?", I had to pull and scan all of it.