Add SqliteMemoryStore.compact_episodic(keep_recent) that prunes only prunable items past a cap

[hadamrd]

Problem

SqliteMemoryStore (src/forge_loop/memory/store.py) exposes put / get / list_active / supersede but no compaction. The memory-generation paths in src/forge_loop/runner/learning.py (record_merged_outcomes, record_failed_outcomes) upsert one durable EPISODIC MemoryItem per issue and never reclaim it, while assemble_boot_context loads every active item. Over hundreds of issues the per-issue failed/merged episodes grow monotonically and bury the load-bearing decisions a maestro needs after a reset.

The event log already solved the mirror problem — eventlog/sqlite.py compaction preserves everything eventlog/models.is_load_bearing accepts and prunes the rest. Curated memory needs the same paired arc.

Concrete example: A repo runs 300 issues. list_active() returns ~3 SEMANTIC decisions, ~5 PROCEDURAL skills, and ~592 EPISODIC failed/merged episodes. Boot context is 99% noise. After store.compact_episodic(keep_recent=50) it returns ~58 items: all 8 load-bearing items survive, plus the 50 most-recent episodes; 542 is the returned prune count.

Dependency

This ticket consumes is_load_bearing_memory(item) from sibling #427. Before implementing, confirm that predicate exists in the memory package (src/forge_loop/memory/models.py or a sibling module) and import it. If #427 has not landed, STOP and surface the blocker rather than duplicating the classifier — the epic requires a single source of truth for load-bearing classification.

Acceptance criteria

• SqliteMemoryStore.compact_episodic(keep_recent: int) -> int exists and returns the number of items pruned.
• Classification uses is_load_bearing_memory (from Add a pure is_load_bearing_memory(item) predicate for curated memory #427) — never an inline re-implementation. Load-bearing items (SEMANTIC, REJECTED_PATH_TAG-tagged, PROCEDURAL) are never deleted regardless of age or count.
• Among active prunable items (superseded_by IS NULL AND is_load_bearing_memory(item) is False, i.e. EPISODIC episodes), the most-recent keep_recent are retained and the older surplus is deleted. "Most-recent" = highest rowid (insertion order), matching list_active's ORDER BY rowid ASC.
• Primary acceptance: given D load-bearing items + M prunable episodic items and keep_recent=k, after compaction all D survive and exactly min(M, k) most-recent prunable items remain.
• Deterministic & idempotent: running twice in a row prunes a non-negative count the first call and exactly 0 the second (already at/under cap is a no-op). Returns 0 when M <= keep_recent.
• keep_recent=0 prunes all prunable items and preserves all load-bearing items. keep_recent < 0 raises ValueError (mirror the validation style in models.py).
• Deletes are wrapped in a with self._connection: transaction (mirror supersede); the prune count derives from the actual delete, not an estimate.

Test matrix

Add to tests/test_memory_store.py (follow the existing _item(...) helper + tmp_path / "memory.db" conventions):

• Unit — primary: D=2 (one SEMANTIC, one PROCEDURAL) + M=5 EPISODIC, keep_recent=2 → returns 3; survivors are both load-bearing items + the 2 newest episodes (assert by memory_id).
• Unit — load-bearing immunity: a REJECTED_PATH_TAG-tagged item plus old EPISODIC items, keep_recent=0 → the rejected-path item survives, all episodes pruned.
• Unit — idempotency: call compact_episodic(2) twice; second call returns 0 and leaves the set unchanged.
• Unit — under cap no-op: M=3, keep_recent=10 → returns 0, nothing deleted.
• Unit — durability: after compaction, reopen the DB (SqliteMemoryStore(db) again) and assert survivors persisted (mirrors test_active_..._round_trip_after_reopen).
• Adversarial / sad-path: keep_recent=-1 raises ValueError; AND a store of only load-bearing items (M=0) → returns 0, deletes nothing even with keep_recent=0.

No integration/e2e changes required for this ticket (compaction is not yet wired into a runner caller — see Out of scope).

Out of scope

• Do not wire compact_episodic into any caller (runner/learning.py, boot, scheduler) — call-site integration is a separate follow-up.
• Do not add the is_load_bearing_memory predicate here — that is Add a pure is_load_bearing_memory(item) predicate for curated memory #427.
• Do not touch the doctor memory_integrity probe — that is Surface prunable-memory backlog in the forge-loop doctor memory_integrity probe #429.
• Do not add the method to the MemoryStore Protocol or FakeMemoryStore unless a test needs it; keep the change to the concrete SqliteMemoryStore. (If you do add it to the Protocol, you MUST also implement it in src/forge_loop/_testing/memory_store.py to keep the contract test green.)
• Do not soft-delete / supersede prunable items — this is a hard DELETE of episodic noise.
• No schema/migration changes; reuse the existing memory_items table and rowid.

File pointers

• src/forge_loop/memory/store.py — add compact_episodic method on SqliteMemoryStore (near supersede). Reuse is_load_bearing_memory, _select, and the with self._connection: pattern.
• src/forge_loop/memory/models.py — source of is_load_bearing_memory (from Add a pure is_load_bearing_memory(item) predicate for curated memory #427); read-only here.
• tests/test_memory_store.py — all new tests.
• Reference mirror (read-only): src/forge_loop/eventlog/sqlite.py + src/forge_loop/eventlog/models.py is_load_bearing for the preserve/prune design.

Verification gate:

uv run pytest tests/test_memory_store.py -q
uv run ruff check src/forge_loop/memory tests/test_memory_store.py
uv run mypy src/forge_loop/memory tests/test_memory_store.py

Original report

Parent: #426

Part of epic: "Bound curated memory with a load-bearing-preserving compaction arc"

Add one method SqliteMemoryStore.compact_episodic(keep_recent: int) that deletes the oldest prunable (per is_load_bearing_memory) memory items beyond keep_recent, retaining the most-recent ones, and NEVER deleting a load-bearing item regardless of age. Deterministic and idempotent (re-running with the same cap is a no-op once at/under the cap). Returns the count pruned. Primary acceptance: given D load-bearing items + M prunable episodic items and keep_recent=k, after compaction all D survive and exactly min(M,k) most-recent prunable items remain. ~100 net LOC incl. tests.

Customer story

A maestro resuming forge-loop after context loss is the customer whose boot context this bounds — the reclaim arc that stops episodic memory from growing monotonically while preserving every load-bearing decision.

[hadamrd]

Blocked on #427 — STOP gate triggered

Per this issue's own dependency clause, I confirmed the prerequisite before implementing and it has not landed:

• is_load_bearing_memory(...) exists nowhere in src/ or tests/ (verified against the whole memory package: __init__.py, models.py, store.py, curator.py).
• Add a pure is_load_bearing_memory(item) predicate for curated memory #427 ("Add a pure is_load_bearing_memory(item) predicate for curated memory") is OPEN with no open/merged PR and no branch (gh pr list --state all --search 427 returns only the unrelated feat(doctor): surface prunable-memory backlog in memory_integrity probe (#429) #430; git log --all has no 427 commit).

The issue body explicitly instructs: "If #427 has not landed, STOP and surface the blocker rather than duplicating the classifier — the epic requires a single source of truth for load-bearing classification." Acceptance criterion #2 further forbids an inline re-implementation. Shipping compact_episodic now would require either (a) inlining the classifier — a direct acceptance-criteria violation and a manifesto Q7 duplicate-capability finding (sev1/sev2, P0 for the critic), or (b) speculatively authoring #427's predicate under this branch, which defeats the epic's single-source-of-truth design and risks diverging from #427's own (unseen) acceptance criteria/test matrix.

Resolution: land #427 first (it defines and exports the shared predicate from the memory package), then this ticket becomes a clean consume-and-import. Holding #428 until #427 merges.

[hadamrd]

forge-loop: worker iteration loop exhausted after max attempts.

• last state: dirty_no_commit
• worktree: /tmp/forge-forge-loop/wt-loop-428
• PR: (none)

The branch is left in place for human inspection. Run cd /tmp/forge-forge-loop/wt-loop-428 && git status to pick up where the loop stopped.

[hadamrd]

forge-loop attempt — ⚠️ NO PR (~122s)

fingerprint: 52233033dc0f

{
  "ts": "2026-06-08T15:03:48+00:00",
  "status": "no_pr",
  "pr_url": null,
  "duration_s": 122.2,
  "note": "",
  "event_count": 0,
  "brief_fingerprint": "52233033dc0f029fa1b02eb04b12003ea644093f5bb5bd6b4246802cb8a8a8d1"
}