v1.0.0 — supersession-aware memory + direct-retrieval reasoning

Continuum v1.0.0

Broke the May-2026 ~34% LongMemEval-S ceiling to 60.8% judged (500 Q) with a
single-shot retrieve→answer pipeline — without the iterative reasoning the
prior report predicted we'd need. The IterativeReasoner was built, A/B-tested,
and cut as net-negative; it ships in-tree as a documented negative result.

Highlights

• Hybrid retrieval — BM25 ⊕ cosine ⊕ Reciprocal Rank Fusion, session-aware.
• LTM + bi-temporal supersession (invalidated_at) — knowledge-update 98.7% recall.
• Direct answer mode — the winning architecture.
• LLM judge as the primary metric (+ offline re-judge), OpenRouter provider,
  SmallLLM, policy engine + 8 policies, token-budget optimizer chain.
• Clean 258-file tree (library + reproduction harness); .env.example for onboarding.

Honest gaps (see CHANGELOG → Known gaps)

• temporal-reasoning 41.4% — a genuine model-reasoning limit; the v1.1 frontier.
• LOCOMO vs Mem0 is preliminary, not a published claim.
• OpenRouter cost accounting unwired.

Verified

All CI green: lint + mypy --strict, tests (3.11 & 3.12), bench + regression gate.
1100+ tests; keyed regression gate locks the shipped pipeline at ≥0.6 judged.

Full notes: see CHANGELOG.md. Next: v1.1 roadmap (on release-2.0).

🤖 Generated with Claude Code