Tool reference · Save & recall
memory_recall_iterative
Answers multi-step questions by splitting them into sub-questions, searching for each, and letting an LLM planner decide when enough evidence has been found.
local stdio server read-only
Caution. Needs an LLM provider for decomposition and planning; each iteration adds at least one LLM call, so it is much slower than memory_recall.
When to use
- The question needs facts from several records chained together ("which service did the person who wrote the auth module move to?").
- A single memory_recall returns only part of what you need.
- You want to see which sub-queries were run and what each one found.
Parameters
| Name | Type | Default | Description |
|---|---|---|---|
queryrequired | string | — | — |
project | string | — | — |
max_iters | integer | 4 | (min 1, max 12) |
k_per_iter | integer | 10 | (min 1, max 50) |
llm_model | string | "configured" | — |
Example
Arguments
{
"query": "Why did we move the billing job off Redis, and what replaced it?",
"project": "my-api",
"max_iters": 3,
"k_per_iter": 5
}Result shape
{
"query": "Why did we move the billing job off Redis, and what replaced it?",
"iterations_used": 2,
"terminated_reason": "converged",
"sub_queries": [
"billing job Redis lock problem",
"billing job lock replacement"
],
"partial_answers": [
"The Redis lock expired mid-run twice."
],
"evidence_count": 2,
"evidence": [
{
"id": 1842,
"type": "decision",
"project": "my-api",
"content": "Use PostgreSQL advisory locks for the nightly billing job…",
"score": 0.91
}
],
"provenance": {
"iters": [],
"elapsed_ms": 1840.2,
"evidence_count": 2
}
}terminated_reason is one of converged, max_iters, no_progress, decomposer_empty, search_error or planner_error.
Values are illustrative; the keys follow the server's handler. MCP clients receive the result as JSON text content.
Server description
The description the server sends to your agent in tools/list, captured from the v14.7.0 source:
IRCoT-style iterative retrieval. Decomposes the query into sub-questions, retrieves per sub-question, and asks a planner LLM whether more retrieval is needed. Best for multi-hop questions. Returns unified evidence + provenance per iteration.