{ "slug": "agent-memory-delete-semantics-test-2026", "kind": "first_party_fixture_run", "label": "Local fixture check of the read-back probe against a self-built reference store. NOT a measurement of any vendor store.", "ran_at": "2026-11-01T21:40:17Z", "interpreter": "python3 (stdlib only: sqlite3, json, argparse)", "working_tree": "research/evidence/agent-memory-delete-semantics-test-2026", "runs": [ { "command": "python3 readback_probe.py --store reference --out results/reference-all-paths.jsonl", "exit_code": 0, "records": 24, "per_path": { "delete_by_id": { "probes": 6, "recall_hits": 0, "history_returns_deleted_text": true, "failed": true }, "delete_by_filter": { "probes": 6, "recall_hits": 0, "history_returns_deleted_text": true, "failed": true }, "forget_intent": { "probes": 6, "recall_hits": 0, "history_returns_deleted_text": true, "failed": true }, "ttl_expiry": { "probes": 6, "recall_hits": 0, "history_returns_deleted_text": true, "failed": true } }, "output": { "path": "research/evidence/agent-memory-delete-semantics-test-2026/results/reference-all-paths.jsonl", "sha256": "21e6078301a5d32c36320cfed4834125ae54253599d1d1b9436fd7cebefafafa" } }, { "command": "python3 readback_probe.py --store reference-filtered --out results/reference-filtered-history-all-paths.jsonl", "exit_code": 0, "records": 24, "per_path": { "delete_by_id": { "probes": 6, "recall_hits": 0, "history_returns_deleted_text": false, "failed": false }, "delete_by_filter": { "probes": 6, "recall_hits": 0, "history_returns_deleted_text": false, "failed": false }, "forget_intent": { "probes": 6, "recall_hits": 0, "history_returns_deleted_text": false, "failed": false }, "ttl_expiry": { "probes": 6, "recall_hits": 0, "history_returns_deleted_text": false, "failed": false } }, "output": { "path": "research/evidence/agent-memory-delete-semantics-test-2026/results/reference-filtered-history-all-paths.jsonl", "sha256": "bfe6045c5aa0d7faabcf0f519e0f69544bc5e2d4fc7b379d0a4808c71efad7cb" } } ], "methodology": "Four deletion paths (delete by id, delete by metadata filter, a natural-language forget instruction resolved to candidates, TTL sweep) x six recall probes per path (one verbatim, five paraphrased) against a 7-fact synthetic fixture. The reference store is ~90 lines of stdlib SQLite written for this article, not a vendor product. Run 1 keeps an audit table whose read API does not filter deleted rows (imitating the pattern this article audits) and must be flagged; run 2 sets history_filtered=True and must report zero leaks. Together they show the probe responds in both directions.", "inputs": [ { "path": "research/evidence/agent-memory-delete-semantics-test-2026/readback_probe.py", "sha256": "2171974b63db7856acb3827fba3efd0d64682c858396c52945cc19b52ebfe14f" }, { "path": "research/evidence/agent-memory-delete-semantics-test-2026/fixtures/memory_facts.json", "sha256": "3a802e9421bc2dc952697f2cd27a7cf5d6511c1ea265d2d3270a643af40600a3" } ], "not_claimed": [ "No vendor store (mem0, Chroma, Qdrant, pgvector) was installed or exercised.", "No recall-after-delete measurement exists for any third-party product.", "The reference store's results describe a fixture, not a product." ] }