memory_reasoning
Shipped in v3-alpha (Beta). Interface may still change before 3.0 stable.
Description
Check if you’ve solved a similar problem before. Returns cached reasoning to save tokens.
Input schema
{
action: "resolve" | "record" | "feedback",
resolve?: { query: string },
record?: { input; reasoning; output; model },
feedback?: { reasoningId; outcome: "success" | "failure" },
}Example
// Before executing — check cache
const cached = await mcp.callTool("memory_reasoning", {
action: "resolve",
resolve: { query: "How to debug Redis connection refused?" },
});
if (cached.hit) {
// Use cached reasoning, suggest cheaper model
// → cached.suggestedModel = "haiku"
}
// After execution — record outcome
await mcp.callTool("memory_reasoning", {
action: "record",
record: { input, reasoning, output, model: "sonnet" },
});
// Later — report whether the cached result actually worked
await mcp.callTool("memory_reasoning", {
action: "feedback",
feedback: { reasoningId: "rsn_abc", outcome: "success" },
});Backed by
- CoT distillation — concept
- Memory Worth — gates which cached reasoning gets served