Skip to main content

What's New

Track the latest features, improvements, and fixes across the entire Cortex ecosystem.

Single Source of Truth

This is the complete changelog for all Cortex packages. Individual package CHANGELOG.md files have been consolidated here for easier tracking.

January 2026​

All Packages v0.33.0 · Jan 20, 2026​

Artifacts API - Interactive Document Management

📦 New Artifacts API for managing interactive, versioned documents
🔄 Full streaming support with pause/resume capabilities
⏪ Version history with undo/redo functionality
🐍 Full Python SDK parity

Show full details

New Features​

Artifacts API

A complete system for managing interactive, versioned digital documents that AI agents can create, modify, and share:

  • Core CRUD Operations: Create, read, update, delete artifacts with full type safety
  • Version History: Track changes with undo(), redo(), getHistory(), and getVersion()
  • Streaming Support: Real-time content streaming with state management (draft → streaming → paused → final)
  • Multi-tenancy: Built-in tenant isolation for SaaS applications
  • File Storage: Support for binary artifacts with Convex storage integration

TypeScript SDK

// Create an artifact
const artifact = await cortex.artifacts.create({
  memorySpaceId: 'user-123-space',
  title: 'Code Snippet',
  content: 'function hello() { return "world"; }',
  kind: 'code',
  tags: ['javascript'],
});

// Version control
await cortex.artifacts.undo(artifact.artifactId);
const history = await cortex.artifacts.getHistory(artifact.artifactId);

Python SDK

# Create an artifact
artifact = await cortex.artifacts.create(
    CreateArtifactOptions(
        memory_space_id="user-123-space",
        title="Meeting Notes",
        content="# Standup Notes\n\n- Discussed roadmap...",
        kind="text",
    )
)

# Streaming
await cortex.artifacts.start_streaming(
    StartStreamingParams(artifact_id=artifact.artifact_id)
)

Vercel AI SDK Integration

New AI tools for artifact management within chat interfaces:

  • createArtifact - Create artifacts via AI
  • updateArtifact - Modify existing artifacts
  • appendToArtifact - Stream content to artifacts
  • React hooks: useArtifacts() and useArtifact()

Changes​

  • Updated @ai-sdk/openai to ^3.0.13
  • Updated @ai-sdk/anthropic to ^3.0.16
  • Updated @types/react to ^19.2.9
  • Updated framer-motion to ^12.27.5

Vercel AI Provider v0.32.0 · Jan 20, 2026​

Reasoning Panel Nativity - Memory Layer Visualization

🧠 New createLayerStreamObserver() helper reduces API route boilerplate from ~40 lines to ~5 lines
⚛️ New useLayerTracking React hook for client-side layer state management
📦 New /react subpath export for React-specific utilities

Show full details

New Features​

Server-Side: createLayerStreamObserver()

A helper function that creates an OrchestrationObserver pre-wired to emit layer events to a Vercel AI SDK stream writer:

import { createLayerStreamObserver, createCortexMemoryAsync } from '@cortexmemory/vercel-ai-provider';
import { createUIMessageStream, createUIMessageStreamResponse } from 'ai';

export async function POST(req: Request) {
  const { observer, emitTo } = createLayerStreamObserver();

  return createUIMessageStreamResponse({
    stream: createUIMessageStream({
      execute: async ({ writer }) => {
        emitTo(writer);
        const cortexMemory = await createCortexMemoryAsync({
          layerObserver: observer,
          // ...config
        });
        // Stream response...
      },
    }),
  });
}

Client-Side: useLayerTracking Hook

A React hook that manages layer tracking state and provides a handleDataPart callback for useChat:

import { useLayerTracking } from '@cortexmemory/vercel-ai-provider/react';
import { useChat } from '@ai-sdk/react';

function ChatComponent() {
  const { layers, isOrchestrating, handleDataPart } = useLayerTracking();

  const { messages, sendMessage } = useChat({
    onData: handleDataPart,
  });

  return (
    <>
      {isOrchestrating && <MemoryLoadingIndicator />}
      <LayerVisualization layers={layers} />
      <Messages messages={messages} />
    </>
  );
}

Event Types

Three transient stream events are emitted during memory orchestration:

  • data-orchestration-start - When orchestration begins
  • data-layer-update - For each layer status change (pending → in_progress → complete)
  • data-orchestration-complete - When all layers finish

Reduction in Boilerplate

ComponentBeforeAfter
API Route layer observer setup~40 lines~5 lines
Client-side data parsing~30 lines1 line (handleDataPart)
Layer state management~100 lines0 (included in hook)
Total~170 lines~6 lines

Quickstart Template Updated

The Vercel AI Quickstart template now uses these new helpers out of the box:

  • app/api/chat-v6/route.ts uses createLayerStreamObserver()
  • lib/layer-tracking.ts re-exports from @cortexmemory/vercel-ai-provider/react

Run cortex update --sync-template to get the updated quickstart files.

CLI v0.32.0 · Jan 20, 2026​

Auto-Discovery of Unregistered Apps

🔍 cortex update now auto-discovers apps in deployment directories
📝 Prompts to register discovered apps for template sync
🚀 Apps created before the registration system are now detected

Show full details

New Features​

Auto-Discovery

The update command now scans deployment directories for Cortex apps that aren't registered in ~/.cortexrc. This helps users who created apps before the .cortex registration system was introduced.

Detected patterns:

  • quickstart/ folders with Next.js + @cortexmemory/vercel-ai-provider
  • Any subdirectory with @cortexmemory/sdk dependencies
$ cortex update --sync-template

   Found 1 unregistered app(s) in deployment directories:
     • my-deployment-quickstart (vercel-ai-quickstart) at /path/to/quickstart

? Register these apps for template sync? › (Y/n)
     ✓ Registered my-deployment-quickstart (vercel-ai-quickstart)

Once registered, apps receive template updates via --sync-template.

All Packages v0.31.1 · Jan 20, 2026​

Dependency Updates

📦 Updated all dependencies to latest versions across the monorepo
🔄 TypeScript SDK, CLI, Vercel AI Provider, and Python SDK

Show full details

Changes​

npm packages updated:

  • @types/node ^25.0.7 → ^25.0.9
  • @typescript-eslint/eslint-plugin ^8.53.0 → ^8.53.1
  • @typescript-eslint/parser ^8.53.0 → ^8.53.1
  • convex ^1.31.4 → ^1.31.5
  • @ai-sdk/openai ^3.0.9 → ^3.0.12
  • @ai-sdk/anthropic ^3.0.12 → ^3.0.15
  • @ai-sdk/react ^3.0.32 → ^3.0.44
  • ai ^6.0.30 → ^6.0.42
  • @hono/node-server ^1.19.8 → ^1.19.9
  • framer-motion ^12.26.1 → ^12.27.1
  • next ^16.1.1 → ^16.1.4

Version bumps:

  • @cortexmemory/sdk 0.31.0 → 0.31.1
  • @cortexmemory/cli 0.31.0 → 0.31.1
  • @cortexmemory/vercel-ai-provider 0.29.0 → 0.29.1
  • cortex-memory (Python) 0.31.0 → 0.31.1

CLI v0.31.0 · Jan 15, 2026​

Adaptive Batch Sizing for cortex db clear

🔄 Automatic batch size reduction on Convex 16MB read limit errors
📉 Starts at 10,000 records, reduces to 500 → 250 → 50 on failures
🧹 Suppresses Convex SDK error noise during retry operations
✅ Reliably clears tables with large embeddings (memories, facts)

Show full details

Bug Fix​

cortex db clear fails on tables with embeddings

The cortex db clear command would fail with "Too many bytes read in a single function execution (limit: 16777216 bytes)" when clearing tables containing records with vector embeddings (memories and facts tables).

Root Cause: Each memory/fact record can contain a 1536-dimension float64 embedding (~12KB per record). With a batch size of 1000 records, this alone approaches the 16MB Convex read limit before accounting for content, metadata, and previousVersions.

Solution: Implemented adaptive batch sizing that:

  1. Starts optimistically with 10,000 records per batch
  2. On Convex "Server Error" (wraps byte limit error), automatically reduces batch size
  3. Retry sequence: 10,000 → 500 → 250 → 50
  4. Continues clearing with smaller batches until complete
  5. Suppresses Convex SDK console.error output during retries for clean CLI output
// New batch size sequence
const BATCH_SIZE_SEQUENCE = [10000, 500, 250, 50] as const;

The fix is resilient to varying record sizes and automatically adapts without requiring manual configuration.

TypeScript SDK v0.31.0 · Jan 14, 2026​

Configurable Recall Limits - Prevent 16MB Read Errors

⚙️ New RecallLimits interface for granular control over recall operations
🔧 Environment variable configuration with per-call overrides
📊 Per-source limits for memories, facts, and graph traversal
🛡️ Eliminates Convex "Too many bytes read" errors in large memory spaces

Show full details

New Features​

Configurable Recall Limits

The recall() API now accepts granular limits to control data retrieval across all sources:

const result = await cortex.memory.recall({
  memorySpaceId: "user-123-space",
  query: "user preferences",
  limits: {
    memories: 20,           // Max vector memories to fetch
    facts: 15,              // Max facts to fetch
    graphHops: 2,           // Max graph traversal depth
    graphEntitiesPerHop: 5, // Entities to expand per hop
    graphResultsPerEntity: 3, // Results per entity from graph
    total: 30,              // Final result cap after merge/rank
  },
});

Configuration Hierarchy

Limits can be configured at three levels (highest priority first):

  1. Per-call overrides - RecallParams.limits
  2. Environment variables - CORTEX_RECALL_*
  3. SDK defaults - Hardcoded sensible defaults

New Environment Variables

VariableDefaultDescription
CORTEX_RECALL_LIMIT_MEMORIES20Max vector memories per recall
CORTEX_RECALL_LIMIT_FACTS15Max facts per recall
CORTEX_RECALL_GRAPH_HOPS2Graph traversal depth
CORTEX_RECALL_GRAPH_ENTITIES_PER_HOP5Entities expanded per hop
CORTEX_RECALL_GRAPH_RESULTS_PER_ENTITY3Results fetched per entity
CORTEX_RECALL_LIMIT_TOTAL30Final result cap

Backward Compatibility

The legacy limit parameter is still supported as an alias for limits.total:

// Legacy (still works)
await cortex.memory.recall({ ..., limit: 25 });

// New (recommended)
await cortex.memory.recall({ ..., limits: { total: 25 } });

Technical Details​

New Types in src/types/index.ts:

export interface RecallLimits {
  memories?: number;
  facts?: number;
  graphHops?: number;
  graphEntitiesPerHop?: number;
  graphResultsPerEntity?: number;
  total?: number;
}

export interface RecallParams {
  // ... existing fields ...
  limits?: RecallLimits;
  limit?: number; // Legacy alias for limits.total
}

New Config Module src/config.ts:

  • RECALL_DEFAULTS - Centralized defaults with env var support
  • resolveRecallLimits() - Merges per-call, env, and defaults

Bug Fix:

Removed dangerous .collect() fallback in Convex queries that could fetch unlimited data. All queries now properly respect limits.

TypeScript SDK v0.30.1 · Jan 14, 2026​

Enriched Entity Extraction & Bidirectional Graph Traceability

🏷️ LLM extraction now returns typed entities (person, organization, place, product, concept)
🔗 New EXTRACTED_WITH edge links Facts to source Memory for full bidirectional traversal
📊 Entity relations (subject-predicate-object triples) sync to graph as typed edges
🔋 Batteries-included: all entity/relation enrichment automatic when graph sync is enabled

Show full details

New Features​

Enriched Entity Extraction

The LLM fact extraction prompt now extracts named entities with semantic types:

// Example extracted fact with enriched entities
{
  fact: "Sarah works at Acme Corp in San Francisco",
  factType: "knowledge",
  confidence: 0.95,
  entities: [
    { name: "Sarah", type: "person" },
    { name: "Acme Corp", type: "organization" },
    { name: "San Francisco", type: "place" }
  ],
  relations: [
    { subject: "Sarah", predicate: "works_at", object: "Acme Corp" },
    { subject: "Acme Corp", predicate: "located_in", object: "San Francisco" }
  ]
}

Bidirectional Fact-Memory Traceability

New EXTRACTED_WITH edge enables traversing from Facts back to their source Memory:

Memory ──[REFERENCES]──► Conversation
   ▲
   │ EXTRACTED_WITH (NEW)
   │
Fact ──[EXTRACTED_FROM]──► Conversation
  │
  └──[MENTIONS]──► Entity ──[works_at]──► Entity

This enables queries like "find all facts extracted from memories about topic X".

Technical Details​

New Types in src/llm/index.ts:

  • ExtractedEntity - Entity with name, type, and optional fullValue
  • ExtractedRelation - Subject-predicate-object triple for graph edges
  • ExtractedFact.entities - Array of extracted entities
  • ExtractedFact.relations - Array of relation triples

Graph Sync Changes:

  • syncFactRelationships() now creates EXTRACTED_WITH edge to Memory node
  • Entity nodes created with semantic type property
  • Relation predicates become typed graph edges (e.g., WORKS_AT, LOCATED_IN)

API Changes:

  • ConflictCandidate interface updated to include entities/relations
  • StoreFactParams already supports entities/relations (no change needed)
  • No breaking changes to public API

TypeScript SDK v0.30.0 · Jan 13, 2026​

Semantic Search for Facts + True Batteries-Included Embeddings

🔍 Native vector/embedding search directly on the facts table
🔋 NEW: Configure embedding once at SDK init - auto-generates everywhere!
🧠 Automatic embedding generation during remember() and recall()
🧪 New extreme multi-turn conversation stress tests

Show full details

New Features​

Batteries-Included Embedding Configuration (v0.30.0+)

Configure embedding provider once - SDK auto-generates embeddings for recall() queries and remember() facts:

// Configure once at SDK init
const cortex = new Cortex({
  convexUrl: process.env.CONVEX_URL!,
  embedding: {
    provider: 'openai',
    apiKey: process.env.OPENAI_API_KEY,
  },
});

// recall() auto-generates embeddings - no manual code!
const result = await cortex.memory.recall({
  memorySpaceId: "user-space",
  query: "What colors does the user like?",
  // Embedding is auto-generated from query!
});

// remember() auto-generates embeddings for facts - no manual code!
await cortex.memory.remember({
  memorySpaceId: "user-space",
  userMessage: "My favorite color is purple",
  agentResponse: "Got it!",
  // Facts are auto-embedded!
});

Or use environment variables for zero-config:

export CORTEX_EMBEDDING=true
export OPENAI_API_KEY=sk-...

Semantic Fact Search

Facts now support native vector search via embeddings:

// Direct: Use semanticSearch() for facts
const facts = await cortex.facts.semanticSearch(memorySpaceId, embedding, {
  minConfidence: 80,
  limit: 20,
});

Extreme Stress Tests

New comprehensive stress test suite (tests/stress/multi-turn-chaos.test.ts):

  • Forgetful User: 50+ repeated questions testing deduplication
  • Indecisive User: 30+ preference changes testing supersession chains
  • Topic Flooder: 100+ similar memories testing semantic search precision
  • Combined Chaos: 100+ turn ultimate stress test
  • Parallel Chaos: 5 concurrent users testing isolation

Technical Details​

Schema Changes

  • Added embedding field to facts table (optional, float64[])
  • Added by_embedding vector index with 1536 dimensions (OpenAI-compatible)

API Changes

  • StoreFactParams.embedding - optional embedding for fact storage
  • UpdateFactInput.embedding - optional embedding for fact updates
  • facts.semanticSearch() - new method for vector-based fact retrieval
  • recall() - automatically uses semantic fact search when embedding available

Integration

  • Vercel AI provider automatically benefits from semantic fact search
  • No changes required to quickstart template

TypeScript SDK v0.29.1 · Jan 10, 2026​

Belief Revision Heuristic Improvements

🧠 Improved decision accuracy when LLM is not configured
🎯 Same subject + same object now correctly returns UPDATE/NONE
🔄 Same subject + different object + related predicate returns SUPERSEDE
➕ Different predicate classes correctly return ADD (no false supersessions)

Show full details

Fixed​

Belief Revision Default Heuristics

When no LLM is configured, the getDefaultDecision() heuristic now correctly handles all 4 decision types:

ScenarioDecisionDescription
Empty memory spaceADDNo conflicts to check
Different subject + low similarityADDUnrelated facts
Different predicate classADDIndependent facts (e.g., "favorite color" vs "favorite food")
High similarity + higher confidenceUPDATERefine existing fact
Same subject + same object + higher confidenceUPDATEConfirm/strengthen existing
Same subject + same object + same/lower confidenceNONEAlready captured
Exact duplicateNONESkip redundant storage
Same subject + different object + related predicateSUPERSEDEPreference changed (e.g., blue → purple)

Key Improvements:

  1. Predicate similarity check - Facts with different predicate classes (e.g., "favorite color" vs "favorite food") no longer incorrectly supersede each other, even when they share the same subject.

  2. Object comparison for same-subject facts - When subject matches:

    • Same object → UPDATE (higher confidence) or NONE (same/lower confidence)
    • Different object + related predicate → SUPERSEDE

  3. Comprehensive test coverage - 27 integration tests now verify all decision paths explicitly.

Technical Details​

  • Added arePredicatesRelated() helper function to check predicate similarity
  • Enhanced getDefaultDecision() to consider subject, object, AND predicate relationships
  • New test file: tests/facts-revision-decisions.test.ts with exhaustive edge case coverage

TypeScript SDK v0.29.0 · Jan 9, 2026​

Automatic Fact and Entity Graph Sync

🔗 Graph sync now automatic when CORTEX_GRAPH_SYNC=true
📊 Facts extracted via remember() automatically sync to graph
🏷️ Entity nodes created with MENTIONS relationships
⚠️ Breaking: syncToGraph option removed from all APIs

Show full details

Major Changes​

Automatic Graph Synchronization

Graph database synchronization for facts and entities is now automatic when CORTEX_GRAPH_SYNC=true:

  • Facts stored via remember(), facts.store(), or facts.revise() automatically sync to graph
  • Entity nodes are created from fact.entities array
  • MENTIONS relationships link Fact nodes to Entity nodes
  • Predicate-based relationships (e.g., WORKS_AT, KNOWS) created from fact.relations
  • SUPERSEDES relationships created when belief revision supersedes facts
  • Graph sync is gated entirely by environment variable

Breaking Change

The syncToGraph option has been removed from all APIs:

// Before (v0.28.x)
await cortex.facts.store(params, { syncToGraph: true });

// After (v0.29.0+) - automatic when CORTEX_GRAPH_SYNC=true
await cortex.facts.store(params);

APIs affected:

  • cortex.facts.store(), update(), delete()
  • cortex.memory.remember(), forget(), delete()
  • cortex.vector.store(), update(), delete()
  • cortex.conversations.create(), addMessage(), delete()
  • cortex.contexts.create(), update(), delete()
  • cortex.memorySpaces.register()

Migration: Remove { syncToGraph: true/false } from all API calls. Set CORTEX_GRAPH_SYNC=true in your environment to enable graph sync.

Technical Details​

  • BeliefRevisionService.executeDecision() now calls syncFactToGraph() and syncFactRelationships() after all fact operations
  • All layer APIs check if (this.graphAdapter) instead of if (options?.syncToGraph && this.graphAdapter)
  • Graph sync is non-blocking - failures are logged but don't fail the main operation

Vercel AI Provider v0.29.0 · Jan 10, 2026​

Automatic Graph Sync Compatibility

🔗 Removed deprecated syncToGraph option from all memory operations
📊 Graph sync now automatic when enableGraphMemory=true with configured adapter
🧪 Updated tests to reflect new automatic sync behavior

Show full details

Breaking Changes​

  • Removed syncToGraph option - Graph sync is now automatic: 
    // Before (v0.28.x)
    await cortexMemory.remember("Hello", "Hi", { syncToGraph: true });

    // After (v0.29.0+) - automatic when enableGraphMemory is configured
    await cortexMemory.remember("Hello", "Hi");

Changed​

  • CortexMemoryProvider.doGenerate() no longer passes syncToGraph to remember()
  • CortexMemoryProvider.doStream() no longer passes syncToGraph to rememberStream()
  • Manual remember() method no longer forwards syncToGraph option
  • ManualRememberOptions.syncToGraph marked as deprecated (ignored if passed)

Fixed​

  • Quickstart template: Fixed neo4j-driver/rxjs bundling issues with Next.js
  • Quickstart template: Added proper webpack externals configuration
  • Quickstart template: Fixed AI SDK v6 type compatibility in memory-agent.ts

CLI v0.29.0 · Jan 10, 2026​

Enhanced Update Display & Graph Template Improvements

📊 Version transitions now show current → latest format
🎨 Color-coded display: green (up-to-date), yellow (outdated)
🔗 Basic template now uses async initialization for graph support
✅ Accurate graph status messages based on actual configuration

Show full details

Changed​

  • Update command version display - cortex update now shows version transitions clearly:

    • Up-to-date packages shown in green: 1.0.0
    • Outdated packages show transition: 1.0.0 → 1.1.0 (yellow → green)
    • Not installed packages shown dimmed (update only affects installed packages)
    • Applies to all package types: SDK, Provider, Convex, Vercel AI
    • Works in both multi-deployment dashboard and single deployment/app views

  • Basic template graph initialization - Now uses Cortex.create() for automatic graph configuration:

    • Async initialization enables auto-configuration from environment variables
    • CORTEX_GRAPH_SYNC=true with NEO4J_URI or MEMGRAPH_URI auto-connects graph adapter
    • Graph sync is automatic on remember() calls (no syncToGraph option needed)

Fixed​

  • Dev-linked apps now refresh on update - In dev mode, apps with file:... references are no longer skipped:

    • Previously, dev-linked apps showed "Everything is up to date!" even when source changed
    • Now always runs npm install to pick up local SDK source changes
    • Fixes scenario: bump local SDK version → cortex update --dev → changes reflected

  • Accurate graph status messages - Template now checks for both flag AND URI:

    • Shows ✓ Graph memory connected (auto-sync active) when fully configured
    • Shows ℹ Graph sync enabled but no database URI configured when URI missing
    • Previously showed misleading "enabled" message even without database URI

Example Output​

● my-deployment
    Path: /path/to/project
    SDK: 0.28.0 → 0.29.0
    Convex: 1.31.0

CLI v0.28.1 · Jan 6, 2026​

Automatic Shell Tab Completion

⌨️ Auto-installs completions for zsh, bash, and fish shells
✨ Dynamic completion for deployment and app names
🧹 Clean removal on uninstall via preuninstall script

Show full details

Added​

  • Automatic shell tab completion - Tab completion auto-installs during npm install -g @cortexmemory/cli:
    • Completes all commands, subcommands, and options with descriptions
    • Dynamic completion of deployment and app names from ~/.cortexrc
    • Completion scripts installed to ~/.cortex/completions/
    • Source line added automatically to shell RC files (idempotent)
    • Clean removal via preuninstall script on npm uninstall -g
    • Manual fallback: cortex completion <zsh|bash|fish> outputs completion script

Changed​

  • Package version display now shows "Vercel AI" instead of "AI" for clarity in cortex update output

TypeScript SDK v0.28.0 · Jan 5, 2026​

Basic Template & Query Performance Fixes

🎯 New headless template with dual CLI/server modes
⚡ Fixed "too many bytes" error in stats queries
🧪 Complete test suite with E2E coverage

Show full details

New Basic Template​

Complete headless demo of Cortex Memory SDK with both CLI and HTTP server modes:

  • Dual-mode operation - Interactive CLI (npm start) or REST API server (npm run server)
  • Optional LLM integration - Works with or without OpenAI API key
  • Rich console output - Animated spinners and memory orchestration visualization
  • Layer observer - Real-time display of all memory layers
  • Full test suite - Unit, integration, and E2E tests included

CLI Commands:

  • /recall <query> - Search memories without storing
  • /facts - List all stored facts
  • /history - Show conversation history
  • /new - Start a new conversation
  • /config - Show current configuration

Fixed​

Query Performance - Resolved "Too many bytes read" error in agents:computeStats:

// Before: Full table scans hitting 16MB limit
const memories = await ctx.db.query("memories").collect();

// After: Indexed queries with sampling
const SAMPLE_LIMIT = 1000;
const memories = await ctx.db
  .query("memories")
  .withIndex("by_participantId", (q) => q.eq("participantId", args.agentId))
  .take(SAMPLE_LIMIT);
  • Uses proper indexes for better performance
  • Limits results to 1000 per query
  • Returns isApproximate: true when sampled

Upgrade: Run npx convex deploy after updating

CLI v0.28.0 · Jan 5, 2026​

Basic Template Tracking & Sessions Support

📁 Basic template projects now tracked in CLI config
🗃️ Sessions and factHistory tables added to db commands
📊 Database stats now cover all 13 tables

Show full details

Added​

  • Basic template tracking in CLI config - Basic template projects are now registered in cortex.config.json:

    • cortex init automatically registers basic projects in the apps section
    • cortex update --sync-template now works with basic template projects
    • cortex config list shows basic projects alongside quickstart apps

  • Sessions and factHistory table support - cortex db clear and cortex db stats now include all Convex tables:

    • sessions - Native session management table
    • factHistory - Belief revision audit trail table
    • Statistics now include counts for all 13 tables

Changed​

  • Added "basic" to AppType union type for template app tracking
  • cortex db clear now clears 13 tables (was 11)

TypeScript SDK v0.27.2 · Jan 1, 2026​

V6 Route Feature Parity Fix

🔧 /api/chat-v6 route now has full feature parity with v5
🧪 Comprehensive E2E tests for quickstart
✅ Fact extraction and belief revision working in v6

Show full details

Fixed​

/api/chat-v6 route now has full feature parity with the v5 route using createCortexMemoryAsync:

  • ✅ Memory recall (pre-call context injection)
  • ✅ Memory storage (post-call conversation saving)
  • ✅ Fact extraction (enableFactExtraction)
  • ✅ Belief revision (superseding outdated facts)
  • ✅ Embedding generation for semantic search
  • ✅ Layer observer for real-time UI updates

Added​

Comprehensive E2E tests for the quickstart covering:

  • Fact storage verification
  • Belief revision (updating preferences through conversation)
  • Memory recall across conversations
  • V5/V6 parity validation
  • Conversation lifecycle (create, list, delete)

TypeScript SDK v0.27.1 · Jan 1, 2026​

AI SDK v6 Agent Architecture Support

🤖 Full integration with Vercel AI SDK v6's ToolLoopAgent
🎯 Type-safe callOptionsSchema for runtime configuration
🔌 createMemoryPrepareCall for automatic memory injection

Show full details

New Exports​

import {
  createCortexCallOptionsSchema,  // Type-safe call options
  CortexCallOptions,
  createMemoryPrepareCall,        // Memory injection via prepareCall
  MemoryInjectionConfig,
  isV6Available,                  // v6 feature detection
  InferAgentUIMessage,            // Type inference for UI messages
} from "@cortexmemory/vercel-ai-provider";

Usage with ToolLoopAgent​

import { ToolLoopAgent } from "ai";

const memoryAgent = new ToolLoopAgent({
  model: "openai/gpt-4o-mini",
  instructions: "You are a helpful assistant with long-term memory.",
  callOptionsSchema: createCortexCallOptionsSchema(),
  prepareCall: createMemoryPrepareCall({
    convexUrl: process.env.CONVEX_URL!,
    maxMemories: 20,
  }),
});

await memoryAgent.generate({
  prompt: "Hello!",
  options: { userId: "u1", memorySpaceId: "app1" },
});

Auto-Detection​

The quickstart automatically detects AI SDK version and routes appropriately:

  • AI SDK v6: Uses /api/chat-v6 with ToolLoopAgent
  • AI SDK v5: Uses /api/chat with streamText

CLI v0.27.3 · Jan 1, 2026​

Neo4j Encrypted URI Scheme Support

🔐 Graph database setup now accepts all neo4j-driver URI schemes
🔒 Support for TLS with system CA validation (+s suffix)
📜 Support for self-signed certificates (+ssc suffix)

Show full details

Added​

  • Neo4j encrypted URI scheme support - All neo4j-driver URI schemes now accepted:
    • bolt://, bolt+s://, bolt+ssc:// (direct connections)
    • neo4j://, neo4j+s://, neo4j+ssc:// (routing/cluster connections)
    • +s suffix for TLS with system CA validation
    • +ssc suffix for TLS with self-signed certificate acceptance

Changed​

  • Updated Docker Compose graph configuration to support optional TLS
  • Added SSL policy configuration for Neo4j bolt connector

December 2025​

Python SDK v0.27.0 · Dec 28, 2025​

Multi-Tenancy & Auth Context System

🔐 Complete multi-tenancy with automatic tenantId propagation
📱 New Sessions API for multi-session management
👤 User profile schemas with validation presets

Show full details

New Auth Module (cortex.auth)​

from cortex.auth import create_auth_context
from cortex import AuthContext, AuthMethod

auth = create_auth_context(
    user_id='user-123',
    tenant_id='tenant-acme',
    organization_id='org-engineering',
    session_id='sess-abc',
    auth_provider='auth0',
    auth_method='oauth',
    claims={'roles': ['admin', 'editor']},
)

cortex = Cortex(CortexConfig(
    convex_url=os.getenv("CONVEX_URL"),
    auth=auth,
))

# All operations automatically scoped to tenant
await cortex.memory.remember(...)
await cortex.conversations.create(...)
await cortex.facts.store(...)

Sessions API​

session = await cortex.sessions.create(CreateSessionParams(
    user_id='user-123',
    tenant_id='tenant-456',
    metadata={'device': 'Chrome on macOS'},
))

await cortex.sessions.touch(session.session_id)
active = await cortex.sessions.get_active('user-123')
await cortex.sessions.end(session.session_id)

User Profile Schemas​

PresetRequired FieldsEmail ValidationMax Size
strictdisplayName, email✓64KB
standarddisplayName✓256KB
minimaldisplayName✗None
noneNone✗None

TypeScript SDK v0.27.0 · Dec 27, 2025​

Multi-Tenancy & Authentication Context

🏢 Complete multi-tenancy support for SaaS platforms
🔐 Automatic tenantId propagation across all APIs
📱 New Sessions API with governance integration

Show full details

AuthContext Integration​

const cortex = new Cortex({
  convexUrl: process.env.CONVEX_URL,
  auth: {
    userId: "user-123",
    tenantId: "tenant-acme",
    sessionId: "sess-abc",
    authMethod: "clerk",
    authenticatedAt: Date.now(),
    claims: { role: "admin" },
  },
});

// All operations automatically scoped to tenant
await cortex.memory.remember({...});
await cortex.conversations.create({...});
await cortex.facts.store({...});

Sessions API​

const session = await cortex.sessions.create({
  userId: "user-123",
  metadata: { device: "mobile", ip: "..." },
});

await cortex.sessions.touch(session.sessionId);
const activeSessions = await cortex.sessions.getActive("user-123");
await cortex.sessions.expireIdle({ maxIdleMs: 30 * 60 * 1000 });

Key Features​

  • ✅ Automatic TenantId Propagation
  • ✅ Sessions API for multi-session management
  • ✅ Auth Validators for format validation
  • ✅ Framework-Agnostic (Auth0, Clerk, NextAuth, Firebase)
  • ✅ Graph Integration with tenant boundaries
  • ✅ GDPR Compatible cascade deletion

CLI v0.27.2 · Dec 28, 2025​

Multi-Deployment Update Command

🔄 cortex update now checks all enabled deployments by default
📊 Color-coded version status table
🎯 Sequential updates with summary

Show full details

Added​

  • Multi-deployment update command - cortex update now checks all enabled deployments:
    • Displays status table with latest SDK/Convex versions and each deployment's current versions
    • Color-coded display (green = up to date, yellow = needs update)
    • Prompts to confirm updating all deployments that need updates
    • Sequential updates with summary at the end
    • -d, --deployment <name> flag for single-deployment mode

CLI v0.27.1 · Dec 27, 2025​

App Lifecycle Management

🛑 Stop command detects and stops running template apps
🔍 Port-based process detection fallback
📊 Enhanced status dashboard with app information

Show full details

Added​

  • App lifecycle management in cortex stop:

    • -a, --app <name> option to stop a specific app
    • --apps-only flag to stop only apps (skip Convex/graph)
    • Apps tracked via PID files (.cortex-app-{name}.pid)

  • Port-based process detection when PID files don't exist:

    • Detects Convex on port 3210 for local deployments
    • Detects apps by their configured port (default 3000)

  • Enhanced cortex status dashboard:

    • Displays running apps with PID and port information
    • Shows detection method (via PID file or via port)

CLI v0.27.0 · Dec 26, 2025​

Vercel AI Quickstart Integration

🚀 Optional demo app installation during cortex init
📱 Template apps management and tracking
⚡ Default enabled for init-created resources

Show full details

Added​

  • Vercel AI Quickstart integration - Optional demo app installation:

    • Installs as /quickstart subfolder
    • Full Next.js app with chat interface and real-time memory visualization
    • Auto-configured with Convex URL and OpenAI API key

  • Template apps management:

    • New apps section in config (~/.cortexrc)
    • Apps shown in cortex config list
    • cortex start automatically starts enabled apps

  • Default enabled - Deployments and apps from cortex init enabled by default

TypeScript SDK v0.26.1 · Dec 26, 2025​

Vercel AI SDK v6.0 Support

✅ Extended peer dependencies to accept ai v6.x
🔧 No breaking changes - fully backward compatible

Show full details

Changed​

  • Extended peerDependencies - @cortexmemory/vercel-ai-provider now accepts ai versions ^3.0.0 || ^4.0.0 || ^5.0.0 || ^6.0.0
  • Users on Vercel AI SDK v6.x will no longer see peer dependency warnings

Create v0.26.0 · Dec 23, 2025​

Cortex CLI Integration

🔧 Optional @cortexmemory/cli installation during setup
📜 Adds CLI scripts to generated package.json
✨ Enhanced success messages with CLI commands

Show full details

Added​

  • Cortex CLI Integration:

    • New optional step to install @cortexmemory/cli during project setup
    • Automatic CLI installation as dev dependency when selected
    • Adds scripts: npm run cortex, cortex:setup, cortex:stats, cortex:spaces

  • User Experience:

    • Updated configuration summary shows CLI installation status
    • Enhanced success message with CLI commands when installed

TypeScript SDK v0.26.0 · Dec 23, 2025​

Enhanced Belief Revision - Subject+FactType Matching

🧠 New pipeline stage catches conflicts missed by pattern matching
🔋 "Batteries included" mode - works without LLM configuration
🔧 Fixed SUPERSEDE and UPDATE actions

Show full details

New Pipeline Stage​

NEW FACT → [Slot Match] → [Semantic Match] → [Subject+Type Match] → [LLM/Heuristic] → Execute
                                                    │
                                            Same subject AND factType?
                                            → Candidate for review

Key Improvements​

  • ✅ Subject+FactType Matching (Stage 2.5) - Catches conflicts with same subject AND factType
  • ✅ Batteries-Included Mode - Works WITHOUT LLM using getDefaultDecision() heuristics
  • ✅ Fixed SUPERSEDE Action - Now uses facts.supersede mutation properly
  • ✅ Fixed UPDATE Action - Uses updateInPlace to avoid creating unwanted versions
// Works WITHOUT LLM configuration
const cortex = new Cortex({ convexUrl: "..." });

await cortex.memory.remember({
  memorySpaceId: "user-space",
  userMessage: "Actually, I prefer purple now",
  agentResponse: "Got it!",
  userId: "user-123",
});
// Old "blue" fact properly SUPERSEDED

Python SDK v0.26.0 · Dec 23, 2025​

OrchestrationObserver API

📊 Real-time monitoring of remember() and remember_stream() pipeline
🐛 Fixed user_id propagation in fact extraction
🧠 Subject+FactType matching for belief revision

Show full details

OrchestrationObserver​

class MyObserver:
    def on_orchestration_start(self, orchestration_id: str) -> None:
        print(f"Starting: {orchestration_id}")

    def on_layer_update(self, event: LayerEvent) -> None:
        print(f"Layer {event.layer}: {event.status} ({event.latency_ms}ms)")

    def on_orchestration_complete(self, summary: OrchestrationSummary) -> None:
        print(f"Done in {summary.total_latency_ms}ms")

result = await cortex.memory.remember(
    RememberParams(..., observer=MyObserver())
)

Bug Fixes​

  • Fixed user_id, participant_id, and source_ref not propagating to facts during belief revision
  • Fixed SUPERSEDE action to use dedicated facts:supersede mutation
  • Fixed UPDATE action to use facts:updateInPlace

CLI v0.26.2 · Dec 25, 2025​

Non-Interactive Convex Setup

🤖 Init wizard sets up Convex without interactive prompts
🔄 Three streamlined setup paths
🔑 Automatic Convex login handling

Show full details

Added​

  • Non-interactive Convex setup in cortex init:

    • Automatically detects Convex authentication status
    • Retrieves team slug automatically from login status
    • Uses Convex CLI flags for seamless setup

  • Three streamlined setup paths:

    • Local development - Cloud project with local backend (recommended)
    • Cloud project - New cloud deployment with full features
    • Existing project - Connect to existing Convex deployment

CLI v0.26.1 · Dec 25, 2025​

Environment Variable Fixes

🔧 Fixed cortex dev overwriting local deployment configs
🔑 Fixed OpenAI API key not saved during init

Show full details

Fixed​

  • cortex dev overwriting .env.local - Inherited CONVEX_* variables were polluting child processes
  • OpenAI API key not saved - Init wizard now correctly saves the configured key

CLI v0.26.0 · Dec 23, 2025​

Secure Password Generation

🔐 Cryptographically secure passwords for Neo4j/Memgraph
🔑 OpenAI API key setup during init

Show full details

Added​

  • Secure password generation - cortex init generates 20-character passwords for graph databases
  • OpenAI API key setup - New optional step with sk- prefix validation

TypeScript SDK v0.24.0 · Dec 20, 2025​

Belief Revision System

🧠 Intelligent fact management preventing duplicates
🔄 Slot-based, semantic, and LLM-based conflict resolution
📜 Complete audit trail with fact history

Show full details

The Problem Solved​

Previously, fact storage was append-only:

  • Conflicting facts accumulated
  • No semantic understanding of when facts should update vs. add
  • No history of how knowledge evolved

Now with Belief Revision​

const result = await cortex.facts.revise({
  memorySpaceId: "user-123-space",
  fact: {
    fact: "User prefers purple",
    subject: "user-123",
    predicate: "favorite color",
    object: "purple",
    confidence: 90,
  },
});

console.log(result.action); // "SUPERSEDE"
console.log(result.reason); // "Color preference has changed"

New API Methods​

MethodPurpose
facts.revise()Full belief revision pipeline
facts.checkConflicts()Preview conflicts without executing
facts.supersede()Manually supersede one fact with another
facts.history()Get change history for a fact
facts.getSupersessionChain()Get lineage of fact versions
facts.getActivitySummary()Analytics on fact changes

Python SDK v0.24.0 · Dec 19, 2025​

Belief Revision System

🧠 Intelligent fact management with conflict resolution
🔄 Pipeline: Slot matching → Semantic → LLM resolution
📜 Fact history and audit trail

Show full details

Usage​

result = await cortex.facts.revise(ReviseParams(
    memory_space_id="agent-1",
    fact=ConflictCandidate(
        fact="User prefers purple",
        subject="user-123",
        predicate="favorite color",
        object="purple",
        confidence=90,
    ),
))

print(f"Action: {result.action}")  # SUPERSEDE
print(f"Reason: {result.reason}")  # "Color preference has changed"

Available Actions​

ActionWhen Used
ADDGenuinely new information
UPDATERefines existing fact
SUPERSEDEReplaces contradictory fact
NONEAlready captured

TypeScript SDK v0.23.0 · Dec 19, 2025​

Unified Context Retrieval with recall()

🔍 The retrieval counterpart to remember()
🎯 Get LLM-ready context from all memory layers
🔗 Graph expansion for discovering related context

Show full details

Usage​

const result = await cortex.memory.recall({
  memorySpaceId: "user-123-space",
  query: "user preferences",
});

// Inject directly into LLM prompt
const response = await llm.chat({
  messages: [
    { role: "system", content: `Context:\n${result.context}` },
    { role: "user", content: userMessage },
  ],
});

Features​

  • ✅ Batteries Included - All sources enabled by default
  • ✅ Graph Expansion - Discovers related context via relationships
  • ✅ Unified Deduplication - Removes duplicates across sources
  • ✅ Multi-Signal Ranking - Semantic similarity, confidence, importance, recency
  • ✅ LLM-Ready Formatting - Structured markdown context

Python SDK v0.23.0 · Dec 19, 2025​

recall() Orchestration API

🔮 Unified context retrieval counterpart to remember()
🎯 Multi-signal ranking with configurable weights
📊 Source breakdown in results

Show full details

Usage​

result = await cortex.memory.recall(
    RecallParams(
        memory_space_id="agent-1",
        query="user preferences",
    )
)

# Use directly in LLM prompts
print(result.context)

Ranking Algorithm​

SignalWeightDescription
Semantic35%Vector similarity score
Confidence20%Fact confidence (0-100)
Importance15%Memory importance (0-100)
Recency15%Time decay (30-day half-life)
Graph Connectivity15%Connected entity count

TypeScript SDK v0.22.0 · Dec 19, 2025​

Cross-Session Fact Deduplication

🎯 Facts no longer duplicated across conversations
🔄 Three deduplication strategies: semantic, structural, exact
⬆️ Confidence-based updates

Show full details

The Problem Solved​

Session 1: "My name is Alice" → Fact created ✅
Session 2: "I'm Alice" → Duplicate detected, skipped ✅
Session 3: "Call me Alice" → Duplicate detected, skipped ✅
Result: 1 fact instead of 3!

Deduplication Strategies​

StrategySpeedAccuracy
semanticSlowerHighest
structuralFastMedium
exactFastestLow

Configuration​

await cortex.memory.remember({
  ...params,
  factDeduplication: "semantic", // Default
});

Python SDK v0.22.0 · Dec 19, 2025​

Cross-Session Fact Deduplication

🎯 Automatic duplicate fact prevention
🔄 Configurable deduplication strategies
📈 Confidence-based updates for higher-quality facts

Show full details

The Solution​

# Deduplication is ON by default
await cortex.memory.remember(
    RememberParams(
        memory_space_id="agent-1",
        conversation_id="conv-123",
        user_message="I'm Alex",
        agent_response="Nice to meet you!",
        user_id="user-123",
        user_name="Alex",
        agent_id="assistant",
        extract_facts=my_fact_extractor,
    )
)

Strategies​

StrategySpeedAccuracy
none⚡ FastestNone
exact⚡ FastLow
structural⚡ FastMedium
semantic🐢 SlowerHigh

November 2025​

Vercel AI Provider v0.2.0 · Nov 24, 2025​

Enhanced Streaming with rememberStream()

🚀 Direct integration with rememberStream() API
📊 Comprehensive streaming metrics
🔄 Progressive fact extraction and graph sync

Show full details

New Features​

  • Progressive Storage - Store partial responses during streaming
  • Streaming Hooks - onChunk, onProgress, onError, onComplete
  • Comprehensive Metrics - First chunk latency, throughput, estimated costs
  • Progressive Fact Extraction - Extract facts incrementally during streaming
  • Progressive Graph Sync - Sync to graph databases during streaming
  • Error Recovery - Resume tokens and partial failure strategies

Configuration​

const cortexMemory = createCortexMemory({
  convexUrl: process.env.CONVEX_URL!,
  memorySpaceId: "my-chat",
  userId: "user-123",
  streamingOptions: {
    storePartialResponse: true,
    progressiveFactExtraction: true,
  },
  streamingHooks: {
    onProgress: (event) => console.log(event),
  },
});

Create v0.2.0 · Nov 24, 2025​

Smart Version Detection

🔍 Auto-fetches latest SDK version from npm
🔄 Dynamic Convex version sync with peer dependencies
✨ Graceful fallback to safe defaults

Show full details

Added​

  • CLI now automatically fetches latest SDK version from npm registry
  • Dynamically detects correct Convex version from SDK's peerDependencies
  • Template always uses "latest" for SDK

Vercel AI Provider v0.1.0 · Nov 5, 2025​

Initial Release

🎉 First release of Cortex Memory Provider for Vercel AI SDK
🔌 Works with OpenAI, Anthropic, Google, Groq
🌐 Edge runtime compatible

Show full details

Core Features​

  • createCortexMemory() - Factory function for memory-augmented models
  • Automatic memory search before each LLM call
  • Automatic memory storage after each response
  • Works with streaming and non-streaming responses

Memory Management​

  • cortexMemory.search() - Manual memory search
  • cortexMemory.remember() - Manual memory storage
  • cortexMemory.getMemories() - Retrieve all memories
  • cortexMemory.clearMemories() - Delete memories

Create v0.1.0 · Nov 2, 2025​

Warning

Deprecated: The create-cortex-memories package has been superseded by @cortexmemory/cli. Use npx @cortexmemory/cli init instead.

Initial Release

🎉 Interactive CLI wizard for Cortex project setup
🐳 Docker integration for Neo4j/Memgraph
📦 Complete project scaffolding

Show full details

Features​

  • npm create cortex-memories → Use npx @cortexmemory/cli init
  • Three Convex setup modes (local/new cloud/existing)
  • Optional graph database integration (Neo4j/Memgraph)
  • Docker detection with platform-specific instructions
  • Automatic dependency installation
  • Backend function deployment

CLI v0.1.0 · Nov 29, 2025​

Initial Release

🎉 First release of @cortexmemory/cli
📋 Complete command suite for memory management
🔧 Multi-deployment support

Show full details

Core Commands​

  • Memory Operations - memory list, search, delete, export, stats
  • User Management - users list, get, delete, export, GDPR cascade deletion
  • Memory Spaces - spaces list, create, delete, archive, participants management
  • Facts Operations - facts list, search, get, delete, export
  • Conversations - conversations list, get, delete, export
  • Convex Management - convex status, deploy, dev, logs, dashboard
  • Database Operations - db stats, clear, backup, restore, export

Features​

  • Configuration management with multiple deployment support
  • Table, JSON, and CSV output formats
  • Interactive confirmations for dangerous operations
  • Dry-run mode for previewing changes

Versioning Policy​

Cortex packages follow semantic versioning:

  • Major (X.0.0): Breaking API changes
  • Minor (0.X.0): New features, backwards compatible
  • Patch (0.0.X): Bug fixes, backwards compatible

Deprecation Policy​

  • Features are marked deprecated for at least one minor version before removal
  • Deprecated features include migration guides
  • Breaking changes are documented with upgrade paths

Roadmap​

Planned for v1.0.0​

  • Complete API stabilization
  • Integration examples for all major frameworks
  • Real-time graph sync worker
  • MCP Server implementation
  • Cloud Mode with Graph-Premium