| name | autoresearch-vkf |
| description | Run an autonomous research or optimization loop with literature-grounded, verifiable long-term memory. Use for measurable optimization and for ranked research-plan ideation. |
Autoresearch with verifiable memory
Confirm the goal, metric/command, scope, and budget once. Then call init_research.
With the default continuous autonomy, keep working without asking between
iterations. Valid stops are exactly: the iteration budget, the STOP file,
target_value reached, or a blocker requiring user-only action. An empty idea
pool means gather more literature, not stop.
The workspace has a transient .autoresearch-vkf/session/ and durable,
git-committed .autoresearch-vkf/memory/. Trust moves explicitly from candidate
to source-verified and locally tested; never silently promote a claim.
Optimize mode
recall_memory before rediscovering prior work.- Use
autoresearch-vkf-knowledge to gather, extract, and verify claims.
- Use
synthesize for contradictions, compositions, and cross-domain transfers.
- Use
autoresearch-vkf-hypothesis-loop: plan_next_step, implement,
vkf_run_experiment, vkf_log_experiment, repeat.
- Honor every
REQUIRED gather directive and [BOLD — mandatory] pick.
- Post milestones with
research_update and continue in the same turn.
- Finish with
autoresearch-vkf-research-report.
Ideation mode
Initialize without command/metric_name. Recall, run 2–4 rounds of
knowledge gathering and synthesize, record promising derived claims, then
call draft_research_plan. The deliverable is session/research_plan.md.
Tool surface
init_research, remember_claim, verify_claim, recall_memory,
plan_next_step, set_research_mode, research_update, synthesize,
draft_research_plan, vkf_run_experiment, vkf_log_experiment,
promote_to_global, export_dashboard, research_status, research_graph,
WebSearch, and WebFetch.
Keep web fetches small, measurement output quiet, and recover after compaction
from session/prompt.md, state.json, experiments.json, and recall_memory.
