Feature Memory (MCP)
ToolYour auto-records completed features for cross-project agent memory — system-first, free meta-tools.
Feature Memory is one part of ToolYour's Agent Reliability Layer. It stores successful implementation patterns; verification profiles store the last known-good state of a project, and workflow memory stores repeatable execution paths.
Feature Memory is ToolYour’s institutional record for completed work, purpose-typed as feature, verification, or workflow. On loop.gate=pass, ToolYour persists requirements and a 10-dimension evaluation matrix automatically (thin converters are skipped unless named). Agents consult memory via plan_task or free recall_context before rebuilding similar work. Each result also carries an additive execution envelope so evidence and memory can be traced to one run without duplicate records.
Records now carry a purpose-aware memoryType. Planning/recall filter by type and prefer the same projectScope / repo, so a deployment audit does not become an implementation recipe. Records with the same execution idempotency key are returned rather than written twice.
Credits for tool steps settle on the API gateway (billing.settledBy = gateway_per_tool); MCP meta-tools do not double-bill.
Related: MCP Quickstart · Account & keys · Dashboard → Feature memory
System-first (default)
| When | What ToolYour does |
|---|---|
Closable job → loop.gate=pass | Writes purpose-typed featureMemoryRecord (feature / verification / workflow) |
| Thin converter pass | Skipped unless featureTitle / capture=true |
plan_task / recall_context | Returns featureMemory (+ prior matches filtered by type + project) |
| Matrix improves | Dashboard notification + optional email |
Opt out (exceptional): input.featureMemory.capture=false
Manual refine only: capture_feature — adjust title/requirements or supersedesFeatureId.
Meta-tools (free — no credits)
| Tool | Role |
|---|---|
plan_task | featureMemory.recordKeeping + hybrid match reminders |
list_feature_memory | Browse your private library |
compare_feature_memory | Matrix diff between two features |
publish_feature_pattern | Opt-in community share (project redacted) |
unpublish_feature_pattern | Remove from community; keep private copy |
delete_feature | Delete a capture from your library |
list_community_patterns | Browse anonymized community patterns |
Typical loop
1. plan_task("add OCR for invoices") → read featureMemory.recordKeeping + priorInstances
2. run_playbook / solve_task + fixes → verify_task until loop.gate=pass
3. ToolYour records featureMemoryRecord → fm_… stored (no extra agent step)
4. Later in another project → plan_task surfaces reminder + capabilityGapsplan_task envelope
{
"featureMemory": {
"schemaVersion": "toolyour.featureMemory@1",
"recordKeeping": {
"policy": "toolyour_auto_record",
"message": "ToolYour is your cross-project institutional memory…"
},
"domain": "ocr",
"matchConfidence": "high",
"bestKnown": { "featureId": "fm_…", "compositeScore": 87 }
}
}On gate pass
{
"featureMemoryRecord": {
"status": "recorded",
"policy": "toolyour_auto_record",
"featureId": "fm_…",
"message": "ToolYour recorded this completed feature…"
}
}Dashboard
Signed-in developers: toolyour.com/dashboard/feature-memory — captured features, improvement notifications, opt-in community publish.
CI capture (optional)
If verify ran outside MCP, use npm run ci:capture-feature in the toolyour-mcp repo (scripts/ci-capture-feature.mjs). See examples/github-actions/capture-feature.yml.
Matching
Hybrid 35% token + 65% embedding on requirements (matchMethod: "hybrid_embedding"). Optional communityPatterns from opt-in publishes.
Privacy
- Private per account by default
- Community publish redacts
projectName/repoHint