V2 Preview

The v1.x line went further than this page originally scoped: the 2026-07 retrieval campaign shipped the general BM25 + dense hybrid retriever (enriched lexical + enriched dense + hop-ordered graph expansion under RRF), k-means clustering and ACT-R activation as offline CLI passes, and the v1.6–v1.7 memory engine (the experiential work-memory tier, composite context surfaces, dependency-aware deletion, the knowledge flywheel). See the Roadmap for the shipped list.

What remains genuinely reserved for V2 — the next substrate generation:

• The typed-atom MemIR substrate — claim atoms, bi-temporal authoritative_claims, derives_from provenance edges. Parked with an explicit revisit trigger (THE-235): it lands with a purpose-built claim-extraction pipeline, and pulls reference-counted derived-artifact GC (upgrading forget’s cascade) with it.
• An episode → authored-claim promotion path — deliberately absent today (the experiential store is a membrane; injected episodes structurally cannot reach authored atoms). If V2 adds promotion, the pre-recorded safety precondition applies: full dependency-closure deletion becomes a write-on gate.
• Cross-store fused ranking — the authored vault and the experiential store federate at query time as separate legs by design. A single fused ranking (and the PIT score calibration it would need) is a V2 question, taken up only with a concrete proposal.

Dark v1.x mechanisms (rerankers, learned sparse, query decomposition, the query class router, convex fusion) are not V2 items — each ships behind a flag with a one-command re-test against the golden-set harness, and re-enters whenever a larger eval set or better serving stack changes its measured verdict.

These are previews of direction, not commitments of schedule. The v1.x tool surface is stable; V2 capabilities will be added additively.