fix: support episode decay evidence snapshots

.dev/roadmap/memory-consolidation/current-progress-and-next-steps.md

@@ -1,7 +1,22 @@
 # Memory Consolidation Current Progress and Next Steps
 
 Status: AI-authored draft. Not yet human-approved.
-Last updated: 2026-05-13 17:09 KST
+Last updated: 2026-05-13 17:55 KST
+
+
+
+## v0.1.153 next-step live dogfood checkpoint
+
+Run directory: `/Users/reddit/.agent-memory/reports/v0.1.153-next-steps-20260513T084528/`.
+
+Results from the requested next-step pass:
+
+- Metadata-rich dogfooding produced a clean dogfood-only fresh epoch: `fresh-epoch-dogfood-only-strict.json` passed with trace coverage `1.0`, empty retrieval ratio `0.0`, no unknown/classified metadata gap, no raw query/trace/content samples, and no mutation.
+- Wider post-v0.1.152/post-v0.1.153 fresh epochs still fail because old rows remain in the epoch window: post-v0.1.153 has `low_epoch_observation_trace_coverage` plus `epoch_empty_retrieval_outcome_metadata_gap_classified`; dominant blocker remains `classified_legacy_missing_outcome`, not an unresolved adapter payload gap.
+- Default ranking remains protected: mixed 50-task shadow eval passed `50/50` with zero baseline regressions, but `active_ranking_policy=conservative_legacy`, `candidate_ranking_policy=graph_reinforced_v1`, `default_retrieval_unchanged=true`, `mutated=false`, and migration still requires the explicit migration command/approval.
+- Reviewed trace-candidate promotion remains narrow only: generation/listing are read-only, ordinary conversation auto-approval is false, raw content is not allowed, and bad apply policy/approval exits without mutation.
+- Broad G4/background apply remains blocked: dogfood epoch G4 preview is read-only/no-mutation with `broad_g4_apply_allowed=false`; required green gates are retrieval ranking, rollback replay, live telemetry reconciliation/fresh epoch, and human-reviewed queue approval.
+- While checking broad G4/decay paths, the live v0.1.153 decay-collapse decision hit an episode evidence snapshot bug (`episodes` use `source_ids_json`, not `evidence_ids_json`). Source now has a regression fix and test; release v0.1.154 is required before relying on live decay-collapse decision over episode candidates.
 
 ## v0.1.153 released runtime checkpoint and next runway
 

.dev/status/current-handoff.md

@@ -1,7 +1,22 @@
 # agent-memory current handoff
 
 Status: AI-authored draft. Not yet human-approved.
-Last updated: 2026-05-13 17:09 KST
+Last updated: 2026-05-13 17:55 KST
+
+
+
+## v0.1.153 next-step live dogfood checkpoint
+
+Run directory: `/Users/reddit/.agent-memory/reports/v0.1.153-next-steps-20260513T084528/`.
+
+Results from the requested next-step pass:
+
+- Metadata-rich dogfooding produced a clean dogfood-only fresh epoch: `fresh-epoch-dogfood-only-strict.json` passed with trace coverage `1.0`, empty retrieval ratio `0.0`, no unknown/classified metadata gap, no raw query/trace/content samples, and no mutation.
+- Wider post-v0.1.152/post-v0.1.153 fresh epochs still fail because old rows remain in the epoch window: post-v0.1.153 has `low_epoch_observation_trace_coverage` plus `epoch_empty_retrieval_outcome_metadata_gap_classified`; dominant blocker remains `classified_legacy_missing_outcome`, not an unresolved adapter payload gap.
+- Default ranking remains protected: mixed 50-task shadow eval passed `50/50` with zero baseline regressions, but `active_ranking_policy=conservative_legacy`, `candidate_ranking_policy=graph_reinforced_v1`, `default_retrieval_unchanged=true`, `mutated=false`, and migration still requires the explicit migration command/approval.
+- Reviewed trace-candidate promotion remains narrow only: generation/listing are read-only, ordinary conversation auto-approval is false, raw content is not allowed, and bad apply policy/approval exits without mutation.
+- Broad G4/background apply remains blocked: dogfood epoch G4 preview is read-only/no-mutation with `broad_g4_apply_allowed=false`; required green gates are retrieval ranking, rollback replay, live telemetry reconciliation/fresh epoch, and human-reviewed queue approval.
+- While checking broad G4/decay paths, the live v0.1.153 decay-collapse decision hit an episode evidence snapshot bug (`episodes` use `source_ids_json`, not `evidence_ids_json`). Source now has a regression fix and test; release v0.1.154 is required before relying on live decay-collapse decision over episode candidates.
 
 ## v0.1.153 released runtime checkpoint
 

.dev/status/next-agent-memory-action.md

@@ -1,7 +1,7 @@
 # agent-memory next action
 
 Status: AI-authored draft. Not yet human-approved.
-Last updated: 2026-05-13 17:09 KST
+Last updated: 2026-05-13 17:55 KST
 
 ## Use this first when the user asks
 

src/agent_memory/api/cli.py

@@ -582,10 +582,12 @@ def _ref_safe_evidence_snapshot(db_path: Path, memory_ref: str) -> dict[str, Any
         }
     memory_type, memory_id = parts
     table_by_type = {"fact": "facts", "procedure": "procedures", "episode": "episodes"}
+    evidence_column_by_type = {"fact": "evidence_ids_json", "procedure": "evidence_ids_json", "episode": "source_ids_json"}
     table_name = table_by_type[memory_type]
+    evidence_column = evidence_column_by_type[memory_type]
     with _open_readonly_sqlite(db_path) as connection:
         row = connection.execute(
-            f"SELECT evidence_ids_json, scope FROM {table_name} WHERE id = ?",
+            f"SELECT {evidence_column} AS evidence_refs_json, scope FROM {table_name} WHERE id = ?",
             (memory_id,),
         ).fetchone()
     relations = list_relations_for_node(db_path, node_ref=memory_ref)
@@ -605,7 +607,7 @@ def _ref_safe_evidence_snapshot(db_path: Path, memory_ref: str) -> dict[str, Any
         "memory_type": memory_type,
         "memory_id": memory_id,
         "exists": True,
-        "evidence_id_count": len(_safe_json_list_from_db(row["evidence_ids_json"])),
+        "evidence_id_count": len(_safe_json_list_from_db(row["evidence_refs_json"])),
         "relation_count": len(relations),
         "scope_present": row["scope"] is not None,
         "content_included": False,

tests/test_cli.py

@@ -10,7 +10,7 @@
 import pytest
 
 from agent_memory.api.cli import main
-from agent_memory.core.curation import approve_fact, create_candidate_fact, supersede_fact
+from agent_memory.core.curation import approve_fact, create_candidate_fact, create_episode, supersede_fact
 from agent_memory.core.ingestion import ingest_source_text
 from agent_memory.core.models import RetrievalTraceEntry
 from agent_memory.core.retrieval import retrieve_memory_packet
@@ -9466,6 +9466,95 @@ def test_dogfood_decay_collapse_preview_reports_stale_weak_evidence_without_muta
     assert "query_preview" not in result.stdout
 
 
+
+def test_dogfood_decay_collapse_preview_handles_episode_source_ids_without_mutation(
+    tmp_path: Path,
+) -> None:
+    db_path = tmp_path / "decay-collapse-episode-preview.db"
+    initialize_database(db_path)
+    source = ingest_source_text(
+        db_path=db_path,
+        source_type="note",
+        content="G5e episode source text and token=SHOULD_NOT_LEAK must not leak.",
+        metadata={"project": "g5e-decay-episode"},
+    )
+    episode = create_episode(
+        db_path=db_path,
+        title="G5e stale episode",
+        summary="Episode evidence should be counted from source ids without reading raw content.",
+        source_ids=[source.id],
+        tags=["g5e", "episode"],
+        importance_score=0.25,
+        scope="project:g5e-decay",
+        status="approved",
+    )
+    record_retrieval_observation(
+        db_path,
+        surface="hermes-pre-llm-hook",
+        query="SHOULD_NOT_LEAK stale episode collapse query",
+        preferred_scope="project:g5e-decay",
+        limit=5,
+        statuses=("approved",),
+        retrieval_trace=[
+            RetrievalTraceEntry(
+                memory_type="episode",
+                memory_id=episode.id,
+                label="stale episode candidate",
+                scope="project:g5e-decay",
+                scope_priority=0,
+                text_match_count=1,
+                rank_value=0.7,
+                total_score=0.7,
+            )
+        ],
+        response_mode="verify_first",
+        metadata={"query_preview": "token=SHOULD_NOT_LEAK", "session_id": "g5e-episode"},
+    )
+
+    before_counts = _table_counts(
+        db_path,
+        ["experience_traces", "retrieval_observations", "memory_activations", "facts", "procedures", "episodes"],
+    )
+    env = {**os.environ, "PYTHONPATH": "src"}
+    result = subprocess.run(
+        [
+            sys.executable,
+            "-m",
+            "agent_memory.api.cli",
+            "dogfood",
+            "decay-collapse-preview",
+            str(db_path),
+            "--limit",
+            "20",
+            "--top",
+            "5",
+            "--frequent-threshold",
+            "3",
+            "--min-decay-score",
+            "0.1",
+        ],
+        cwd=Path(__file__).resolve().parents[1],
+        env=env,
+        capture_output=True,
+        text=True,
+    )
+
+    assert result.returncode == 0, result.stderr
+    payload = json.loads(result.stdout)
+    assert payload["read_only"] is True
+    assert payload["mutated"] is False
+    candidates = {candidate["memory_ref"]: candidate for candidate in payload["decay_collapse_candidates"]}
+    candidate = candidates[f"episode:{episode.id}"]
+    assert candidate["ref_safe_evidence"]["evidence_id_count"] == 1
+    assert candidate["ref_safe_evidence"]["content_included"] is False
+    assert _table_counts(
+        db_path,
+        ["experience_traces", "retrieval_observations", "memory_activations", "facts", "procedures", "episodes"],
+    ) == before_counts
+    assert "SHOULD_NOT_LEAK" not in result.stdout
+    assert "source text" not in result.stdout
+
+
 def test_dogfood_supersession_preview_reports_claim_conflicts_without_mutation(
     tmp_path: Path,
 ) -> None: