scottrbk Gathering context before work
A single query answers a question; a context gather prepares a task. The output is a pack the work can lean on, with every claim cited by memory ID and every dry angle stated — silence is never ambiguous.
Invoking operations
Operations are named by registry name (query_memory, get_document, ...). Invoke via
whichever surface this agent has:
- MCP:
execute_forgetful_tool(tool_name="query_memory", arguments={...}) - CLI:
forgetful call query_memory --args '{"query": "..."}' --json
Get any operation's schema at runtime: how_to_use_forgetful_tool (MCP) or
forgetful tools info <operation> (CLI) — schemas are deliberately not repeated here.
Delegating this workflow to a subagent: hand it the CLI flavor — Bash travels into
subagents, MCP tool access does not. Run MCP-flavored gathers in the main conversation.
Step 1 — Derive the query angles
From the task description, write 3–5 distinct angles: the feature area, the technology or library, exact error text or identifiers, the architectural layer, adjacent domains that share patterns. One angle is a spot-check; several are coverage.
Done when: the angles are listed before any query runs.
Step 2 — Recall per angle
query_memory per angle — query and query_context are both required, and
query_context should carry the task's intent (per forgetful-recall: reads cross-project
by default, identifiers verbatim, truncated handled). Skills referenced by name in this
skill resolve via search_skills → export_skill when not already available. Track which
angles hit and which came up dry.
Done when: every angle has run and its outcome is recorded.
Step 3 — Explore around the strongest hits
Where hits reference entities or cluster into a topic, walk the graph (per
forgetful-explore, medium depth): expand the memories, discover the entities, harvest
what is attached to them.
Done when: the strongest hits have their neighbourhood mapped, not just their own text.
Step 4 — Open the linked material
Hits point outward — follow the pointers: get_document for linked documents,
get_code_artifact for reusable code, search_skills + get_skill for procedures that
already do the task, get_skill_links for what those procedures bundle.
Done when: load-bearing linked material is read, not just listed.
Step 5 — Sweep for unlinked material
Link-following only surfaces what a hit already points to — a document or code artifact
that no surfaced memory links to stays invisible to that flow. Close the gap:
list_documents and list_code_artifacts, scoped with project_id when the task is
project-bound, then scan titles for relevance. search_skills already searches skills
directly, so no separate sweep is needed there.
Done when: documents and code artifacts have been swept by title, not just reached by link.
Step 6 — Synthesize the pack
Assemble, in order of usefulness to the task:
- Decisions that constrain the work (cited by memory ID)
- Patterns and preferences that shape how it should be done
- Constraints — things ruled out, with the reason
- Code pointers — artifacts and source references worth opening
- Procedures — stored skills that cover part of the task
- Gaps — every angle that came up dry, stated as "no existing knowledge about X"
Done when: the pack cites its sources, states its gaps, and the task could start from it without re-querying.