This page explains the product and engineering roadmap for moving NoteConnection from a knowledge-visualization system into a verifiable local-first learning platform.
The roadmap is driven by one primary target:
- maximize measurable mastery improvement while keeping retrieval and tutoring outputs explainable and auditable.
This requires:
- stable knowledge representation and evidence traceability,
- explainable retrieval and policy-controlled orchestration,
- mastery-state update loops with testable quality gates.
- Visualization is useful for navigation but does not prove learning effect.
- LLM output without evidence alignment can improve fluency while degrading trust.
- Long-term learning outcomes require both temporal validity and memory governance, not only single-turn responses.
- Unified ingest and staleness rebuild pipeline.
- Relation and temporal graph contracts.
- Store backend abstraction (
file/memory/graphdb) with fallback safety. - Query backend comparability (
local_hybridvskeyword_only) with trend telemetry.
- Mastery diagnostics and misconception tracking.
- Dual path outputs (
MasteryPath+DivergencePath) in session planning. - Session history analytics with strategy-source and outcome telemetry.
- Quality trend and threshold checks linked to path strategy behavior.
- Pluggable tutor actions under evidence-first constraints.
- Memory policy diagnostics and trend governance for
session/unit/long_term. - Runtime capability runbook with remediation queue and incident verification chain.
- Core interfaces are available in
src/learning/api.tsand exported viasrc/learning/index.ts. - Type contracts for atom/evidence/relation/temporal/mastery/action/trace are defined in
src/learning/types.ts. - API surface is wired in
src/server.tsand validated bysrc/knowledge.api.contract.test.ts. - Learning Workbench orchestration and diagnostics are integrated in
src/frontend/path_app.js.
- The branch now contains real Phase-3 tutor/memory slices, but current HEAD should not be described as "Phase-1 closed".
- The accurate state is:
- Phase-1 A8 has advanced into an embedded local-backend baseline: graph/store ops semantics, embedded SQLite graphdb persistence/query paths, and HTTP adapter paths exist, the default runtime now targets
graphdb/sqlitewith explicit file fallback, restart durability is integration-proved, host-level dist/runtime + packaged sidecar proof is in place, and a host-level workload matrix is in place acrosssmoke/medium/heavy; soak / longer-duration / performance hardening are still open. - Phase-1 A9 has advanced into an ANN connector operational baseline: ANN-style prefilter, representation telemetry, remote index sync, live
external_httpquery proof, host-level dist/runtime + packaged sidecar proof, and a host-level workload matrix acrosssmoke/medium/heavynow exist, but recall/latency threshold convergence and release-grade calibration are still open. - Phase-2 now has an operational diagnostics baseline:
learning quality,session plan quality, query comparison, staleness, query-backend config, and query-backend diagnostics are live inKnowledgeLearningPlatform.ts, but they are not yet release-closed because they still require release-grade calibration on top of the current graph/ANN operational baseline. - Phase-3 is now operational-baseline rather than catalog-only: tutor telemetry, tutor trace/provider trends, conversation memory, memory-policy diagnostics, and default runtime tutor-adapter injection are real, but production-proven multi-provider routing policy is still open.
- Phase-1 A8 has advanced into an embedded local-backend baseline: graph/store ops semantics, embedded SQLite graphdb persistence/query paths, and HTTP adapter paths exist, the default runtime now targets
- Active rollout focus therefore changes from "assume closure and move on" to "keep the new A8 packaged/runtime and workload-matrix proofs green, keep the new A9 runtime and workload-matrix proofs green, finish the remaining soak/performance A8 closure plus A9 threshold convergence, then move next into honest Phase-2 gate promotion."
- The current branch now also has a real Tauri-first assistant-reply baseline:
- grounded conversation no longer stops at plain-text
assistantMessagemounting, - backward-compatible
assistantBlocksnow exist in the conversation response contract, - a shared Reader-derived markdown/math/mermaid runtime now powers rich assistant replies in the Tauri workspace shell.
- grounded conversation no longer stops at plain-text
- This does not change the strategic Phase-1 / Phase-2 / Phase-3 classification.
It changes the interaction baseline:
- Tauri is now the rich-render reference surface,
- Godot remains a later downgrade/materialization target rather than the feature-constraining baseline for assistant replies.
- Remote CI has also been re-stabilized at current HEAD:
Migration GatesandFixrisk Operational Readinesswere brought back to green after the strict PathBridge verifier was realigned with current brightness bounds,- workflow Node.js baselines were upgraded from 20 to 24 so CI no longer relies on the forced-node24 compatibility override path.
Practical implication:
- the next work is no longer "make the Tauri shell capable of rich replies" because that baseline now exists,
- the next work is to broaden rich assistant block coverage, keep the render substrate verified in real browser/Tauri runtime flows, and continue reducing architecture pressure while the graph/ANN substrate work proceeds.
- Real graph backend closure:
- keep the new embedded
graphdb/sqlitedefault alive across packaged/runtime paths, - treat restart durability, host-level packaged/runtime proof, and the new
smoke/medium/heavyworkload matrix as already proved, then extend verification into ops-preferred query semantics, fallback consistency, and soak/performance durability against that local graph engine.
- keep the new embedded
- Real ANN connector closure:
- keep the new live
external_httpconnector path healthy under real sync/query traffic, - keep the new host-level ANN runtime and workload-matrix proofs green,
- benchmark recall/latency thresholds before calling the vector layer production-ready.
- keep the new live
- Phase-2 quality gating:
- keep the new
learning quality,session plan quality, query-comparison, and staleness diagnostics aligned with the same runtime truth, - promote those trend diagnostics from observability to release-blocking governance only after Phase-1 backend closure is no longer
Partial+.
- keep the new
- Tutor-routing hardening:
- keep the newly active default
tutorAdapterobservable, - extend from local-first routing into a production-proven multi-provider policy while preserving explicit fallback behavior.
- keep the newly active default
- Architecture pressure:
- continue reducing the major monoliths (
server.ts,KnowledgeLearningPlatform.ts,path_app.js,app.js,routes/knowledge.ts) so roadmap claims and code structure stop drifting apart.
- continue reducing the major monoliths (
- Truth before closure:
- a route, card, or adapter surface is not "done" merely because the contract exists; placeholder methods and catalog-only wiring must stay classified as open work.
- Foundation recovery before rollout claims:
- finish real graphdb/ANN delivery before calling the backbone production-grade.
- Evidence before promotion:
- every new ability must ship with contract wiring, runtime observability, and fresh test evidence.
- Quality gates over demos:
- trend outputs are useful, but release decisions must be tied to non-placeholder threshold-gate outcomes.
- Tutor routing must be active, not only enumerable:
- multi-adapter tutor support is only closed after normal server runtime emits real adapter telemetry under live execution.
Use this page as the execution view (layer matrix, debug runbook, and priority backlog).