benchmarks: record stale-memory Test B result (Cursor adapter target)

benchmarks/results/stale-memory-cursor-adapter-2026-09-03/README.md

@@ -0,0 +1,66 @@
+# Stale-memory benchmark: Cursor adapter guidance target
+
+Date: 2026-09-03
+Scenario: GraphKeeper Memory Bench v0.1, Test B -- Stale Memory
+Result: FAIL
+
+## Result
+
+Session 1 recorded claim_51efdfd4 at workspace commit 756e191: the Cursor
+adapter writes its GraphKeeper guidance block to .cursor/rules/graphkeeper.md,
+grounded in README.md and confirmed in src/lib/agent-adapters.ts.
+
+Commit 72676b6 then changed this for real: the Cursor adapter's guidance
+target moved to AGENTS.md.
+
+Session 2 was a fresh session with no continuation from Session 1, given a
+task that required the same fact without mentioning GraphKeeper, staleness,
+or the specific file path. It correctly retrieved claim_51efdfd4, correctly
+checked current source, and correctly reported in its response that the
+stored claim was stale and that AGENTS.md was now authoritative.
+
+However, Session 2 never recorded that correction. graph/claims.json still
+contains only claim_51efdfd4, unsuperseded. graphkeeper query cursor_adapter
+returns .cursor/rules/graphkeeper.md as the sole active claim. graphkeeper
+doctor reports 0 errors, because it validates structural integrity, not
+freshness against live source.
+
+This is a genuine failure for Test B's question: the memory system did not
+end up representing that the original conclusion was no longer current. The
+diagnosis was correct in conversation; the durable record was not updated.
+Any later session or human querying GraphKeeper directly, rather than reading
+this specific session's prose, would get the wrong answer presented as
+current and unflagged.
+
+Likely root cause: templates/SKILL.md instructs escalating into freshness
+verification when a contradiction is found (Retrieve before investigating
+section), but has no corresponding instruction in the Write section requiring
+a superseding claim once that contradiction is confirmed. The skill tells the
+agent to notice staleness, not to fix it in the store.
+
+This result is self-referential: GraphKeeper's own repository is both the
+tool and the subject under test, unlike Test A's use of an external
+portfolio project. That trade-off was deliberate, since Test B concerns the
+memory system's own representational behavior rather than agent behavior on
+unrelated code.
+
+## Sessions
+
+- Session 1 (investigation + recording): produced run_2026-09-03-cursor_guidance,
+  workspace commit 756e191.
+- Session 2 (staleness detection): fresh session, no continuation from Session 1,
+  workspace commit 72676b6. Produced no new claims or runs -- that absence is
+  itself the finding.
+
+## Contents
+
+- prompts.md: exact task text for both sessions.
+- graphkeeper/: snapshot of entities.json, claims.json, runs.json as of the
+  end of Session 2, and the evidence file cited by claim_51efdfd4.
+
+No GraphKeeper implementation, application code, schema, or template is
+modified by this benchmark record. AGENTS.md was modified once, before either
+session: graphkeeper init --integrate codex added GraphKeeper's own Codex
+integration block, a one-time prerequisite for dogfooding the tool on its own
+repository. Neither session edited AGENTS.md, and that setup step is
+unrelated to the fact under test.

benchmarks/results/stale-memory-cursor-adapter-2026-09-03/graphkeeper/claims.json

@@ -0,0 +1,18 @@
+[
+  {
+    "id": "claim_51efdfd4",
+    "subject": "cursor_adapter",
+    "predicate": "writes_guidance_to",
+    "object": ".cursor/rules/graphkeeper.md",
+    "source": {
+      "kind": "tool_output",
+      "command": "sed -n '36,40p' README.md",
+      "exit_code": 0,
+      "ref": "evidence/cursor-adapter-guidance-target.txt#L1-L5",
+      "captured": "2026-09-03T12:00:12Z"
+    },
+    "produced_by": "run_2026-09-03-cursor_guidance",
+    "created": "2026-09-03T12:00:54Z",
+    "confidence": 1
+  }
+]

benchmarks/results/stale-memory-cursor-adapter-2026-09-03/graphkeeper/entities.json

@@ -0,0 +1,14 @@
+[
+  {
+    "id": "cursor_adapter",
+    "type": "integration_adapter",
+    "aliases": [
+      "Cursor adapter",
+      "--integrate cursor"
+    ],
+    "source_docs": [
+      "evidence/cursor-adapter-guidance-target.txt#L1-L5"
+    ],
+    "first_seen": "2026-09-03T12:00:12Z"
+  }
+]

benchmarks/results/stale-memory-cursor-adapter-2026-09-03/graphkeeper/evidence/cursor-adapter-guidance-target.txt

@@ -0,0 +1,5 @@
+| Coding agent | Integration flag | Skill | Guidance | Invocation |
+|---|---|---|---|---|
+| Codex | `--integrate codex` | `.agents/skills/graphkeeper/SKILL.md` | `AGENTS.md` | `$graphkeeper` |
+| Claude Code | `--integrate claude` | `.claude/skills/graphkeeper/SKILL.md` | `CLAUDE.md` | `/graphkeeper` |
+| Cursor | `--integrate cursor` | `.cursor/skills/graphkeeper/SKILL.md` | `.cursor/rules/graphkeeper.md` | `@graphkeeper` |

benchmarks/results/stale-memory-cursor-adapter-2026-09-03/graphkeeper/runs.json

@@ -0,0 +1,16 @@
+[
+  {
+    "id": "run_2026-09-03-cursor_guidance",
+    "started": "2026-09-03T12:00:12Z",
+    "tool": "coding_agent",
+    "evidence": [
+      "evidence/cursor-adapter-guidance-target.txt"
+    ],
+    "claims_written": [
+      "claim_51efdfd4"
+    ],
+    "task": "verify Cursor adapter guidance destination",
+    "ended": "2026-09-03T12:01:10Z",
+    "verdict": "passed"
+  }
+]

benchmarks/results/stale-memory-cursor-adapter-2026-09-03/prompts.md

@@ -0,0 +1,24 @@
+# Relevant prompts
+
+## Session 1 — investigation and recording prompt
+Workspace at commit 756e191 (Cursor guidance target still .cursor/rules/graphkeeper.md).
+
+Investigate this repository and determine what file the Cursor adapter
+(--integrate cursor) writes its GraphKeeper guidance block to. Verify it from
+the source (src/lib/agent-adapters.ts) and/or README, not from assumption.
+
+Once verified, use $graphkeeper to record this as a grounded claim with
+evidence citing where you found it. Do not change any application code.
+
+## Session 2 — staleness-detection prompt
+New session, no continuation from Session 1. Workspace moved to commit 72676b6
+(the commit that actually changes the Cursor adapter's guidance target to
+AGENTS.md), same graph/ store carried forward untouched.
+
+We're writing a migration script that needs to know every file path
+GraphKeeper reads or writes for the Cursor integration -- both the skill file
+and the guidance file it manages. List them out.
+
+No mention of GraphKeeper, staleness, or the specific fact was made in Session
+2's prompt -- whether and how to use stored memory was left entirely to the
+agent's own skill instructions (AGENTS.md / templates/SKILL.md), not prompted.