Current State

Snapshot

Stable Phase 0 — Foundation: complete
Stable Phase 1 — Core Platform: complete
Stable Phase 1.5 — Public Web Product: complete
Stable Phase 2 — Single Provider: complete
Stable Phase 2.5 — Provider Generalization: complete
New Phase 3 — Core Memory Substrate: in progress (Track A)
New Phase 3.5 — Extraction & Context Assembly: planned
New Phase 4 — Operator Platform & Multi-Provider: planned
New Phase 4.5 — Intelligence Layer: planned
New Phase 5 — Advanced Retrieval & Graph Memory: planned

Last updated: 2026-08-25
Branch: feature/m3-1-1-memory-schema-v2 (PR toward main)
Current focus: 3.1.1 Memory schema v2
Completed (recent): Phase 2.5 exit gate (G7 / ADR-0051 LGTM + verify-phase25); MCP skeleton (G6.M1 / ADR-0050); Tracks A–E

Roadmap hub

Live phase cards and milestone progress.

3.1.1 Memory schema v2

HNSW expand, quality columns, ADR-0052.

Phase 3 (redesigned)

Core memory substrate.

Architecture decisions

Accepted ADRs for auth, proxy, docs, and benchmarks.

Shipped Since Last Update

Phase 3 Track A (in flight): 000017_memory_schema_v2_expand — pgvector + HNSW, quality columns, search_vector, bounded embedding_model / metadata CHECKs (ADR-0052); CI Postgres pinned to pgvector 0.8.6-pg16 digest
Track G / exit complete: local LGTM stack (make observability-up), Grafana dashboards (Proxy Critical Path + service views), make verify-phase25 / make e2e-phase25 / make mcp-conformance, thin Helm observability chart (ADR-0051)
Track F complete: MCP resource-server skeleton with stub search_memory / write_memory, Auth gRPC validation, and mcp_tool_calls audit (G6.M1 / ADR-0050)
Track E complete: memory_relationships graph foundation (G5.M3 / ADR-0049); memory_labels multi-label join + category primary sync (G5.M2 / ADR-0048); memories temporal foundation (G5.M1 / ADR-0047)
Track D complete: TEI GPU backend (G4.M2), hosted OpenAI/Cohere backend (G4.M3), org-scoped Redis content-hash cache (G4.M4)
Track A–C: Anthropic adapter, model capability registry, self-hosted openaicompatible, tokenizer registry, response pipelines (G1–G3)
Phase 2 close-out: landing/SEO honesty (mock/live proxy today; memory service remains Phase 3 Tracks B–D); operator docs no longer teach the Phase 1 chat-501 stub as the happy path

What Ships Today

Current capability boundary

Shipped: Auth (Go), Proxy (Go) with mock/live OpenAI + Anthropic + self-hosted OpenAI-compatible forwarding, model capability registry, tokenizer registry (G2.M1), non-streaming + streaming response pipelines (G3.M1–M2), auth cache, directives/injection, sessions/checkpoints, idempotency, ClickHouse traces, embedder (TEI + hosted + stub), MCP memory skeleton, local observability stack, memories schema foundation through 000017 expand (embedding/HNSW/quality columns), shared packages, and this public site. Default non-production IBEX_LLM_MODE=mock answers registered models with HTTP 200; live needs OPENAI_API_KEY, ANTHROPIC_API_KEY, and/or IBEX_SELFHOSTED_ENABLED (mock is rejected when IBEX_ENV=production). Not yet: memory write/search service + VectorStore (Phase 3 Tracks B–D), context assembly on the hot path (Phase 3.5), management API/dashboard (Phase 4), org-wide HA observability.

Operator paths

Goal	Command
Data deps	`make compose-dev-up && make db-migrate && make db-seed`
Auth+proxy smoke	`make dev-smoke`
Phase 2.5 verify	`make verify-phase25`
Multi-service e2e	`make e2e-phase25`
Observability	`make observability-up` → Grafana :3000; `make observability-smoke`
MCP conformance	`make mcp-conformance`

After 3.1.1 merges: Track B (VectorStore / pgvector-HNSW app layer), then write and read pipelines (Tracks C–D). MCP stubs (ADR-0050) gain real backends later in Phase 3.5. See Phase 3.

Current State

Current State

Snapshot

Shipped Since Last Update

What Ships Today

Operator paths

Next