Skip to main content

Changelog

Historical positioning

Older entries may describe experimental or legacy memory-layer features such as Cortex, eval gates, legacy security automation, or exact MCP tool counts. Treat those entries as release history, not current product positioning. Current Novyx positioning is AI agent change control and blast-radius gating for production actions.

Phases 1-5 — April 10, 2026​

Novyx Control: governance shipment

A five-phase upgrade to Novyx Control that turns governance into a first-class, tenant-authored, per-agent capability. The marketing site, SDKs, MCP server, and docs are all updated.

  • Phase 1 — Policy-as-code. Author custom Control policies in YAML or JSON. Each rule has a regex match, severity, optional context_requires, and an optional on_violation outcome (block, require_approval, or warn). Severity-based defaults if omitted: CRITICAL→block, HIGH→require_approval, MEDIUM/LOW→warn. Tier-gated: Free=0, Starter=5, Pro=25, Enterprise=unlimited. Built-in FinancialSafetyPolicy and DataExfiltrationPolicy are always active. Full CRUD: POST/GET/PUT/DELETE /v1/control/policies. See Custom Policies.
  • Phase 2 — Approval workflows. New action status pending_review. Real approval queue at GET /v1/approvals (returns latest event per action_id ordered by sequence_number, not timestamp — survives cross-worker clock drift). POST /v1/approvals/{action_id}/decision with 404 for unknown action and 409 for already-decided. Three approval modes: Solo (phrase + 5s delay), Team (different person OR 10min cooldown), Enterprise (multi-person chains with min_approvals). See Approval Workflows.
  • Phase 3 — Multi-provider neutrality. novyx-agent 1.x → 2.0: provider and model are now required keyword arguments on Agent(...). The OpenAI default has been removed. Valid providers: openai, anthropic, litellm (with litellm covering Gemini, Mistral, Cohere, Ollama, etc.). Same change applies to nx.create_agent() (Python SDK 3.3.0), nx.createAgent() (JS SDK 3.1.0), and the create_agent MCP tool. New ToolDef.to_provider_schema(provider) dispatcher. Breaking change. See the novyx-agent 2.0 upgrade guide.
  • Phase 4 — Governance dashboard. GET /v1/control/dashboard?window=24h|7d|30d&bucket=hour|day returns aggregated stats — totals, violations broken down by policy and agent, and a time-series. Postgres-only; tenants in file mode receive an empty-but-valid shape with backend: "file". New GET /v1/control/agents/{agent_id}/violations for per-agent violation history. Tier: Starter+. See Governance Dashboard.
  • Phase 5 — Agent-scoped policies. The same policy name can have a tenant-wide version and per-agent overrides. Policy registry cache key is now (tenant_id, agent_id_or_None). When both exist, agent-scoped wins for that agent only. All five policy CRUD endpoints accept optional agent_id. Tier: Pro+. See Agent-Scoped Policies.
  • Phase 6 — SDK wrappers. Both SDKs ship typed helpers for the new endpoints: Python nx.create_policy, nx.list_policies, nx.get_policy, nx.update_policy, nx.delete_policy, nx.governance_dashboard, nx.agent_violations. JS equivalents in camelCase. All policy methods accept optional agent_id. Initially shipped in novyx 3.3.0 (Python) / 3.1.0 (JS).
  • Phase 6.5 — Main-API action helper + bug fixes. Two issues found during the docs verification pass and fixed in 3.4.0 / 3.2.0:
    • nx.submit_action(action, params, agent_id=None) (Python) and nx.submitAction(action, params, { agent_id }) (JS) — typed wrappers around POST /v1/actions for the main cloud governance flow. Distinct from the legacy nx.action_submit(connector, operation, payload) which targets a separate Control instance via control_url. Most users should use submit_action.
    • agent_violations since/until fix. The endpoint previously returned novyx_ram.v1.control.violations_failed when timestamp filters were passed (asyncpg parameter binding issue — Python str types went over the wire as text and the Postgres ::timestamptz cast silently failed). Fix parses to datetime in Python before binding, drops the cast. Bad input now returns 400 novyx_ram.v1.control.invalid_violations_param with the offending field name.

Versions: novyx (Python) 3.4.0 · novyx (JS) 3.2.0 · novyx-agent 2.0.0 · novyx-mcp 2.9.0 line. Current MCP registry: 102 tools.

v3.0.1 — March 11, 2026​

Security Hardening, MCP 91 Tools & JS SDK Parity

  • MCP Server v2.2.0: Expanded from 23 to 91 tools — broad Core API coverage including eval, cortex, replay, actions, drafts, knowledge graph, and legacy security automation
  • JS/TS SDK v2.11.0: 27 new methods for full API parity with Python SDK
  • Security audit fixes: SSRF DNS resolution hardened, credential leak prevention, TTL bypass fixed, search filter push-down for Postgres
  • Control integration: Governed actions, approval flows, policy engine wired into Core audit trail
  • Eval gate: GitHub Action for CI/CD memory health quality gates (marketplace-ready)
  • Cortex enhancements: Real-time contradiction detection, explain_action, webhook bus
  • Event bus: Real-time SSE streams for memory events
  • Revenue loop: Payment failure handling, value preview on tier limit 403s, /v1/usage/insights endpoint, first-memory activation tracking

v3.0.0 — March 8, 2026​

Eval System, AsyncNovyx & Experimental Memory-Health Gates

  • Eval system: Experimental memory health scoring with composite score (0–100) from recall consistency, drift, conflicts, and staleness
  • Eval gate: POST /v1/eval/gate was introduced as an experimental memory-health threshold check. In the current backend, eval is disabled by default unless NOVYX_ENABLE_EXPERIMENTAL=1 is set.
  • Baseline checks: Save recall queries as baselines for experimental memory-health runs
  • Drift analysis: Track how memory composition changes over time — count deltas, importance shifts, topic churn
  • AsyncNovyx: Full async client with httpx — pip install novyx[async]. Complete parity with sync client
  • Eval history: Track health scores over time with per-tier retention (Free: 7d, Starter: 30d, Pro: 90d, Enterprise: 365d)
  • 7 new SDK methods: eval_run(), eval_gate(), eval_history(), eval_drift(), eval_baseline_create(), eval_baselines(), eval_baseline_delete()

v2.11.0 — March 5, 2026​

Compensation Webhooks, Memory Export & Slack/Discord Formatting

  • Compensations API (Pro+): preview, list, get, and acknowledge rollback compensation plans
  • New webhook event: rollback.compensations — fires when a rollback detects ACTION trace steps
  • Memory Export: export memories as Markdown, JSON, or CSV via GET /v1/memories/export
  • Slack and Discord native webhook formatting
  • GitHub Action for CI/CD memory sync and rollback validation

v2.10.0 — March 3, 2026​

Pro Tier Expansion, Milestones API & LlamaIndex

  • Cortex Insights and Insight Config moved from Enterprise to Pro+
  • Replay Counterfactual Recall and Drift Analysis moved from Enterprise to Pro+
  • Milestones API (Pro+): tag meaningful points in your memory timeline
  • novyx-llamaindex v1.0.0 — LlamaIndex integration
  • Webhooks (Pro+): real-time HMAC-signed notifications
  • Teams & RBAC (Pro+): multi-tenant collaboration

v2.9.2 — February 27, 2026​

SDK 2.9.2 — Context Spaces, Replay & Cortex Methods

  • 13 new SDK methods across Python and JS/TS
  • Context Spaces: create_space(), list_spaces(), space_memories(), update_space(), delete_space(), share_space()
  • Replay: replay_timeline(), replay_snapshot(), replay_lifecycle(), replay_diff()
  • Cortex: cortex_status(), cortex_run(), cortex_insights()
  • Full API parity between Python and JS/TS SDKs

MCP v2.0.0 — February 26, 2026​

MCP Server 2.0 — Tools, Resources & Prompts

  • Initial MCP release with core memory, context spaces, replay, and cortex tools
  • 6 resources: novyx://memories, novyx://stats, novyx://usage, novyx://spaces, and per-ID lookups
  • 3 prompts: memory-context, session-summary, space-context
  • Context spaces for multi-agent collaboration via MCP
  • Graceful tier gating — tools return clear upgrade messages on lower tiers