Changelog
What's new in Novyx Core.
v2.4.0Mar 30, 2026— Runtime v2 — Agent Orchestration
- •Novyx Core now manages the full agent lifecycle. Runtime v2 introduces five new primitives that turn Novyx from a memory layer into an orchestration platform.
- •Agents — First-class agent identities that persist across sessions. Register an agent once, track everything it does.
- •Missions — Goal-oriented task assignments with pause, resume, and cancel. Know what your agents are working on and why.
- •Capability Packs — Define which tools an agent can use and at what risk level. Tool governance at the orchestration layer.
- •Checkpoints — Save agent state mid-mission. Roll back to any checkpoint if something goes wrong.
- •Supervisor Interventions — Human-in-the-loop controls. Approve, reject, pause, escalate, or reroute agent decisions in real-time.
- •16 new MCP tools (91 → 107), 26 new API endpoints, full Python + TypeScript SDK support. Tier availability: Agents & Missions (all tiers), Capability Packs (Starter+), Checkpoints (Pro+), Interventions (Enterprise).
novyx-agent v1.1.0Mar 25, 2026— Tiered Memory Recall — Signal-Strength Formatting
- •Recalled memories now formatted with signal-strength tiers: [strong], [moderate], [background]
- •Background memories capped at 2 to reduce noise in agent context
- •Addresses the Karpathy observation: LLMs overfit to low-relevance memories in context
- •Both sync Agent and AsyncAgent use the same formatter
- •98 tests, zero warnings
Site v2.0Mar 25, 2026— Homepage Restructure, Design System & Accessibility
- •Homepage restructured from 14 sections to 6 — focused narrative with Demo CTA elevated to position 4
- •DESIGN.md created — formal design system documenting color tokens, CTA hierarchy, card hover tiers, heading rules
- •OKLCH color system — migrated from HSL/hex to perceptually uniform OKLCH color space
- •Tailwind CSS v4 migration with @theme-based design tokens
- •Mobile pricing: stacked tier cards replace horizontal-scroll table on mobile
- •Accessibility: prefers-reduced-motion support, aria-live on error messages, 44px touch targets
- •Framer Motion animations with scroll-triggered reveals across all pages
- •Native <dialog> modals replacing custom implementations
- •Replay dashboard with timeline scrubber, event list, and memory snapshots
- •Fixed React hydration errors (#418/#423) on all prerendered pages
MCP v2.3.0Mar 22, 2026— MCP Server — Sentinel Intel Migration & Bug Fixes
- •novyx-mcp v2.3.0 published to PyPI
- •Sentinel stores migrated from JSONL to Postgres for multi-machine reliability
- •Replay timeline now records memory operations (create, update, delete, rollback)
- •Fixed 3 Sentry issues: async Postgres datetime handling, sentinel test updates, CI stability
- •Cache invalidation added for sentinel stores across Fly.io machines
MCP v2.2.0Mar 17, 2026— MCP Server — 91 Tools, Full Core Coverage
- •Expanded novyx-mcp from 23 to 91 tools — every Core API endpoint now accessible via MCP
- •New tool categories: eval, control, webhooks, teams, milestones, export, anomalies
- •Revenue loop: payment failure handler, value preview 403s, usage insights, first memory celebration
Security AuditMar 15, 2026— Security Hardening — 3 Audit Rounds
- •Resolved all critical, high, and medium findings across 3 security audit rounds
- •SSRF protection: full DNS resolution validation on all outbound requests
- •Credential leak prevention: stripped sensitive fields from error responses
- •TTL bypass fix: time-to-live enforcement on per-ID lookups
- •Search filter push-down to Postgres (was filtering in Python)
- •Eval router auth fix + approval SDK methods added
- •JS SDK 2.10.0: added 27 missing methods for full API parity with Python
v3.0.1Mar 14, 2026— Tier Restructure — All Features on All Tiers
- •Every feature now available on every tier — pay for scale, not access
- •Memory Drafts: stage changes before committing (25/250/unlimited by tier)
- •Branches: isolated memory namespaces for experimentation (3/25/unlimited by tier)
- •Novyx Control integrated with Core — audit logging, tier checks, SDK + MCP tools
- •Event bus with real-time SSE streams for memory changes
- •Recall explain: see why each memory matched your query
- •Eval gate: block deploys when memory health drops below threshold
Agent SDK v1.0.0Mar 11, 2026— Novyx Agent SDK — Agent Framework with Built-in Persistent Memory
- •novyx-agent v1.0.0 published to PyPI — standalone agent framework with built-in Novyx memory
- •Auto-recall: agent retrieves relevant memories before every response
- •Auto-remember: agent stores important context after every response
- •@tool decorator for defining custom tool functions
- •Multi-provider LLM support: OpenAI, Anthropic, and any LiteLLM-compatible model
- •All Novyx memory features built in — rollback, audit, semantic search, knowledge graph
Hygiene v2.0.0Mar 10, 2026— Novyx Hygiene — Automatic Context Persistence for Claude Code
- •novyx-hygiene v2.0.0 published to PyPI — automatic context persistence for Claude Code
- •PreCompact hook auto-saves session state (task, decisions, files, git state) before /compact
- •SessionStart hooks auto-inject context after compact, clear, or resume
- •Context health scoring: freshness, file sprawl, decision tracking, mixed concerns (A-F grades)
- •CLI commands: hygiene save, resume, inject, score, list, config
- •Optional cloud sync via Novyx API key — semantic search across sessions, cross-machine sync
- •Writes .claude/hygiene.md so Claude reads context automatically between sessions
v3.0.0Mar 8, 2026— Eval System, AsyncNovyx & CI/CD Quality Gates
- •Eval system: memory health scoring with composite score (0-100) from recall consistency, drift, conflicts, and staleness
- •CI/CD quality gate: POST /v1/eval/gate blocks deploys when memory health drops below your threshold (Pro+)
- •Baseline regression testing: save recall queries as baselines, detect when results degrade across eval 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: 7 days, Starter: 30, Pro: 90, Enterprise: 365)
- •7 new SDK methods: eval_run(), eval_gate(), eval_history(), eval_drift(), eval_baseline_create(), eval_baselines(), eval_baseline_delete()
v2.11.0Mar 5, 2026— Compensation Webhooks, Memory Export & Slack/Discord Formatting
- •Compensations API (Pro+): preview, list, get, and acknowledge rollback compensation plans — reversal actions for side effects detected in trace steps
- •New webhook event: rollback.compensations — fires with the full compensation plan in reverse order 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 — set format to "slack" or "discord" and payloads are automatically transformed into Block Kit or embed format
- •GitHub Action for CI/CD memory sync and rollback validation
v2.10.0Mar 3, 2026— Pro Tier Expansion, Milestones API & LlamaIndex
- •Cortex Insights and Insight Config moved from Enterprise to Pro+ — all Pro users now get pattern detection
- •Replay Counterfactual Recall and Drift Analysis moved from Enterprise to Pro+ — full time-travel suite on Pro
- •Milestones API (Pro+): tag meaningful points in your memory timeline (deployments, model swaps, prompt changes) and diff memory state before vs. after
- •SDK: milestone_create(), milestones(), milestone_diff() in Python; milestoneCreate(), milestones(), milestoneDiff() in JS/TS
- •novyx-llamaindex v1.0.0 — LlamaIndex integration with NovyxChatStore, NovyxMemory, and RAG retriever
- •Webhooks (Pro+): real-time HMAC-signed notifications on memory create, update, delete, and rollback with automatic retry
- •Teams & RBAC (Pro+): multi-tenant collaboration with owner, admin, member, and read-only roles
v2.9.2Feb 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 v2.0.0Feb 26, 2026— MCP Server 2.0 — 23 Tools, Resources & Prompts
- •23 MCP tools: core memory (10), context spaces (6), replay (4), cortex (3)
- •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
- •Replay and Cortex tools now accessible via MCP, not just raw API
- •Graceful tier gating — tools return clear upgrade messages on lower tiers
API FixesFeb 26, 2026— API Improvements — LWW, Tag Filtering & Limits
- •Last-write-wins (LWW) conflict resolution now available on all tiers (was Starter+)
- •Tag AND filtering: search memories matching all specified tags
- •PATCH /v1/memories/{id} tag limit increased from 10 to 25
v2.8.0Feb 25, 2026— Cortex — Autonomous Memory Intelligence
- •Cortex API — autonomous memory maintenance for AI agents (Pro+)
- •Consolidation: auto-merge near-duplicate memories using pgvector similarity
- •Reinforcement: boost frequently-recalled memories, decay forgotten noise
- •Insight Generation: pattern detection from memory corpus (Pro+)
- •5 new API endpoints: /v1/cortex/status, config (GET/PATCH), run, insights
- •Configurable per-tenant: thresholds, decay age, feature toggles
- •New /docs/cortex page with full endpoint documentation, Python + JS examples
- •Cortex section added to homepage and pricing page
- •SDK 2.8.0 — cortex_status(), cortex_config(), cortex_configure(), cortex_run(), cortex_insights()
v1.6.0Feb 19, 2026— Replay — Time-Travel Debugging
- •Replay API — memory-level observability for AI agents (Pro+)
- •Memory Timeline: paginated feed of every create, update, delete, and rollback
- •Point-in-Time Snapshots: reconstruct full memory + knowledge graph state at any timestamp
- •Memory Lifecycle: track a single memory from creation to deletion with all events
- •Memory Diff: compare memory state between two timestamps — added, removed, modified
- •Counterfactual Recall (Pro+): re-run semantic search against historical memory state
- •Drift Analysis (Pro+): track tag frequency shifts, importance changes, emerging/disappearing topics
- •New /docs/replay page with full endpoint documentation, Python + JS examples
- •Replay section added to homepage and pricing page
- •SDK 2.7.0 — replay_timeline(), replay_snapshot(), replay_memory(), replay_diff(), replay_recall(), replay_drift()
v1.5.0Feb 19, 2026— Integrations — LangChain, CrewAI, MCP & OpenClaw
- •novyx-langchain — LangChain memory backend with persistent semantic memory, rollback, and LangGraph checkpointer
- •novyx-crewai v1.0.0 — CrewAI storage backend for all 3 memory types (short-term, long-term, entity)
- •novyx-mcp v1.0.0 — MCP server for Claude Desktop, Cursor, and Claude Code with 10 tools
- •New /integrations page with install instructions, quickstart snippets, and PyPI links for all packages
- •OpenClaw section moved from docs to /integrations — 4 extensions with persistent memory tools
- •Templates section: Customer Support Agent, Research Knowledge Agent, Trading Audit Agent
- •Integrations added to navbar and footer navigation
v1.4.0Feb 19, 2026— Knowledge Graph, JS/TS SDK & API Expansion
- •Knowledge Graph API — store structured relationships as subject–predicate–object triples (Pro+)
- •Entity traversal and filtering with auto-dedup and cascade deletes
- •JavaScript / TypeScript SDK published to npm — full API parity with Python, zero dependencies, Node.js 18+ and browsers
- •New JS/TS docs page at /docs/js with complete examples
- •Session-scoped memories — nx.session() groups memories by conversation
- •Memory Links — explicit directional relationships between memories with relation labels and weights
- •New endpoints: GET /v1/memories/{id}, PATCH /v1/memories/{id}, GET /v1/context/now
- •Search API expanded: recency_weight, match_confidence, superseded_by, tag filtering
- •Store API expanded: importance, context, auto_link, confidence fields
- •Python/JS SDK cards on homepage now link to PyPI and npm
- •Hero code snippet upgraded to tabbed Python/JS toggle
v1.3.0Feb 14, 2026— 4-Tier Pricing, Execution Traces & Feature Gating Updates
- •New 4-tier pricing: Free ($0), Starter ($12/mo), Pro ($39/mo), Enterprise ($199/mo)
- •Pro price reduced from $49 to $39/mo
- •Enterprise price reduced from $299 to $199/mo
- •Free tier rollbacks increased from 3/month to 10/month
- •Execution Traces now live on Pro+ — RSA-4096 signed, hash-chained, independently verifiable
- •Context Spaces (formerly Memory Namespaces) shipped on all tiers
- •Memory Space Sharing live on Pro+
- •Starter tier unlocks: all conflict resolution strategies, audit export, AI rollback suggestions
- •Conflict resolution enforcement: Free tier now correctly restricted to reject-only strategy
- •Circuit Breaker confirmed free on all tiers (safety is not a premium feature)
- •Anomaly Alerts live on Pro+
- •353 tests passing, 80 new E2E tests across all tier × feature combinations
v1.2.0Feb 8, 2026— Pricing Simplification & Novyx Core Launch
- •Simplified pricing: Free, Pro ($49/mo), Enterprise ($299/mo)
- •Consolidated to three tiers (Free/Pro/Enterprise)
- •Pro tier pricing reduced from $99 to $49
- •Updated positioning: "Persistent memory + rollback for AI agents"
- •Added Production Trust metrics and Integrations showcase
v1.1.0Feb 2, 2026— Launch
- •Novyx Core public launch
- •SDK v0.1.2 published to PyPI
- •Self-service API key generation
- •Tier enforcement (Free/Pro/Enterprise)
- •Magic Rollback with tier gating (3/month free, unlimited pro/enterprise)
- •Sentinel security API with CORS
- •Status page with live health checks
v1.0.0Jan 2026— Beta
- •Initial RAM API deployment
- •Sentinel threat detection
- •Cryptographic audit trail
- •SDK beta