/query-memory

name: query-memory
description: Recall what is already known — your own prior reasoning, the swarm's shared findings (including dead ends and verifier feedback), and the verified facts — before doing new work. Use when prior conclusions, examples, dead branches, verification outcomes, or verified results may inform the current question, claim, subgoal, or branch decision.

Query Memory

Before spending effort, check what already exists. There are three places to look, in the three-memory model:

1. **Your own local memory** (private): read/grep `local_memory/notes.jsonl` and `events.jsonl` for your prior reasoning and what you already tried. 2. **Global memory** (shared findings): `gm_search(query, kinds=...)` over the swarm's findings. Especially useful kinds:

• `dead_end` / `obstacle` — paths that already died (skip them);
• `verification` — outcomes of others' `fact_submit` (learn from rejections);
• `conclusion` / `example` / `counterexample` / `plan` — others' results to build on.

You can also read the `global_memory/<kind>.jsonl` files directly. 3. **Fact graph** (verified truth): `fact_search(query)` (BM25 over the verified facts) to find results you can cite or that show your subgoal is already proved — it returns `{fact_id, statement}`; read the full proof from `fact_graph/facts/<fact_id>.md` on a relevant hit, and `fact_graph/glossary.json` to reuse the project's symbol definitions. A proof may build **only** on facts (cite a `fact_id`).

Procedure

1. Obey the current prompt's restrictions first. If it forbids a direction, file, or search, that overrides default recall. If it recommends specific results or directions, raise their priority. 2. Start with the cheapest relevant source: your own local memory for your context; `gm_search` for the swarm's findings; the fact graph for verified building blocks. 3. Prefer a narrow, targeted query (specific `kinds`, a sharp query string) over reading everything. 4. **Workspace boundary:** stay inside your own working directory and the shared project stores. Do not scan parent directories, other workers' private `local_memory/`, or other projects.

Retrieval priority

• A relevant **verified fact** (fact graph) is the strongest hit — you can build

on it directly by citing its `fact_id`.

• A sibling's **`dead_end`/`obstacle`** saves you from re-walking a dead path.
• A sibling's **`verification`** rejection tells you why a similar claim failed.
• A `conclusion`/`example`/`counterexample` is awareness — useful, but **never a

brick** (only facts are). Re-verify anything you intend to build on.

Output

Note what you recalled and how you used it in your local memory (`events`). Do not re-publish others' findings; just use them.

Tools

• `gm_search` (recall shared findings; BM25 over global memory)
• `fact_search` (recall verified facts; BM25 over the fact graph — novelty + citation lookup)
• local memory is read directly (no tool — read/grep the files); read a fact's full

proof from its `fact_graph/facts/<fact_id>.md` once `fact_search` surfaces it