▸ WHAT WE ARE BUILDING NEXT

ROADMAP.

Public phases, honest scoping, no vaporware promises. If a phase is marked exploring, there is a real chance it never ships. If it is marked active, the code is being written right now.

  1. Phase 0 — Agent OS core

    ✓ SHIPPEDshipped Q4 2025

    The foundation: PGE trinity loop, six memory tiers, 145 MCP tools, 18 channel adapters, skill registry, gatekeeper with 4-class risk model.

    • ▸Planner → Gatekeeper → Executor loop
    • ▸6-tier memory architecture (chat, episode, vault, identity, tactical, entity)
    • ▸145 MCP tools across 14 modules
    • ▸18 channel adapters (Telegram, Discord, Slack, WhatsApp, Signal, Voice, ...)
    • ▸Built-in skill library (MIT)

  2. Phase 1 — Flutter Command Center

    ✓ SHIPPEDshipped Q1 2026

    Rebuilt the UI layer in Flutter — 33 screens, Material 3 dark theme, i18n across en/de/zh/ar. Replaces the legacy React UI.

    • ▸33 Flutter screens covering every daily-use surface
    • ▸WebSocket chat, voice, vision, push notifications
    • ▸Mobile pairing via HMAC-signed QR
    • ▸Multi-interface networking (Tailscale / ZeroTier / WireGuard / LAN)

  3. Phase 2 — Evolution engine + autonomous learning

    ✓ SHIPPEDshipped Q1 2026

    The assistant that improves itself while you sleep. Idle-gated, budget-capped, checkpoint-interruptible.

    • ▸System detector (CPU / RAM / battery / active windows)
    • ▸Autonomous Learner reviews last 72 hours of runs
    • ▸Meta-Learner scores strategy efficiency over time
    • ▸Budget caps: seconds / tokens / cost / energy
    • ▸Checkpoint + resume across idle windows

  4. Phase 3 — Creator marketplace

    ▶ ACTIVEQ4 2026

    Third-party developers can publish and sell their own agent packs. 70/30 revenue share, no exclusivity, Ed25519-signed pack verification, one-time purchases through Gumroad.

    • ▸Community pack validator (5-check pipeline: syntax, prompt-injection, tool allowlist, safety, signature)
    • ▸Publisher identity + trust levels
    • ▸Pack registry with install / search / report tools
    • ▸Creator dashboard with sales + payout data
    • ▸Reddit Lead Hunter as the first case study

  5. Phase 4 — Windowed reasoning

    ? EXPLORINGexploring

    Research track for long-horizon reasoning on a local model — essentially, how to make a 27B planner work on tasks that need 100k tokens of context without blowing up the KV cache.

    • ▸Contextual summarization via the Episode tier
    • ▸Chunked context with sliding-window promotion
    • ▸ARC-AGI-3 integration as a reasoning testbed
    • ▸Offline CNN training for keyboard-driven games

  6. Phase 5 — Federated learning

    ? EXPLORINGexploring

    Opt-in federation where multiple Cognithor instances share anonymous learning signals without sharing data. Still in brainstorming — there is a real chance this never ships if we cannot prove it can be done without compromising the local-first commitment.

    • ▸Differential privacy on strategy signals
    • ▸User consent per signal class
    • ▸No raw data ever leaves the machine, ever
─── Want to see a specific thing on this roadmap? Open an issue on GitHub or drop it in the blog comments. ───