Roadmap
Shipped (current: v1.34.0)
Section titled “Shipped (current: v1.34.0)”The complete G2.1 tool surface — every tool across all domains — plus everything the v1.x line added on top of the v1.0 hardening gate (OpenTelemetry, Prometheus, CloudEvents, rate limiter, 8-triple native prebuilds, this docs site):
-
The measured retrieval engine (v1.4–v1.7): enriched BM25 + enriched dense + hop-ordered wikilink expansion fused under RRF k=10 — a general hybrid retriever, gated by an n=250 golden set with a statistical ship rule (permutation test, FDR, a non-inferiority floor). Contextual chunk enrichment measured +0.223 nDCG and defaults on. k-means clustering and ACT-R activation recompute run as offline CLI passes. Per-vault GraphRAG edge isolation shipped. Mechanisms that did not earn a default ship dark behind flags. Cross-encoder rerankers and the class router have A/B results on the record (
docs/EVALUATION.md, ADR 0007); learned sparse, ColBERT, convex fusion, query decomposition and MMR have no recorded result (docs/DARK_MECHANISMS.mdis the inventory). Graph densification (derivedshared_tag/ vec0-kNN / LLM Pass-3 edges beyond authored wikilinks,retrieval.densify.*) is wired but unmeasured — it ships dark pending a multi-hop golden-set A/B (the earlier virtual-hop experiment hit an 80% bridge-recall ceiling; the champion figure it was compared against has since been withdrawn as unreproducible, so that comparison needs re-deriving before it means anything). -
The experiential work-memory tier (v1.6): a quarantined second store with serve-path retrieval logging, auto-captured work episodes (poison-scanned, evaluator-stamped eligibility), and reader tools under a strict contract.
-
Composite context surfaces (v1.7):
vault_context(budget-packed one-call context with lesson surfacing and a TTL-enforced prewarm cache) andreflect(grounded synthesis + adversarial challenge + a typed-delta preference profile). -
Dependency-aware deletion with a hash-chained forget audit; Obsidian Git and Remotely Save companion bridges; a knowledge-flywheel CLI family (
metrics,gaps,prefetch,reflect,forget,citation-infer,contribution-report,activation-recompute,cluster); and a vec0 index with a per-vault partition key, rebuilt in place from stored embeddings. -
CycloneDX SBOMs, generated per package on release (
npm sbom) and uploaded as build artifacts (non-blocking). -
A reference page per tool, generated from the live tool registry at docs build time (scopes, annotations, confirmation behaviour, parameter tables, output summary, related tools) and linked from the tool catalog.
Themes under active work, in no particular order — see the tool reference and configuration docs for what already ships today:
- Zero-config semantic search. Lexical search and every note tool already work with nothing but a vault path; closing the gap so semantic and graph-seeded retrieval work out of the box too, without standing up a separate embeddings backend first.
- Remote access with OAuth. The HTTP transport is opt-in today, gated by signed-JWT auth; an OAuth authorization-server flow is in design so a remote agent can authenticate without a hand-provisioned token.
- Broader client coverage. Verifying and documenting behavior across more MCP clients, beyond the ones already covered in the compatibility matrix.
- Memory you own, in Markdown. The experiential store already lives under the same ACL as the rest of the vault; ongoing work makes what it remembers, and why, more legible and more directly editable as plain notes.
Deferred
Section titled “Deferred”- A per-scope HITL-raise (today per-scope overrides only tighten; isolate unattended automation on a second instance instead).
Out of scope for the v1.x line
Section titled “Out of scope for the v1.x line”- The typed-atom MemIR substrate (claim atoms, bi-temporal
authoritative_claims) — a downstream engine-build phase, parked with an explicit revisit trigger. - Cross-store fused ranking between the authored vault and the experiential store — the stores federate at query time by design (separate legs, never one ranked list), so score calibration across them has no substrate (this was investigated and closed with that verdict).