What's new.
Every shipped change, versioned. Breaking changes flagged. RSS feed at /changelog.rss .
All
Features
Breaking
Fixes
Improvements
Security
Forecast strategy + Brain Glow
Feature 9th Dream Engine strategy Forecast — consolidated memories project forward with calibrated confidence intervals on every horizon. Feature Brain Glow score tracking at /brain-glow Improvement Forecast outputs surface in /v1/memory/ask responses under the new forecast key.
Multi-agent namespaces
Feature RBAC-scoped namespaces for agent swarms. Each agent gets an isolated memory context; read and write permissions scoped per role. Feature Pub/sub channels for cross-agent sync. Subscribe any namespace to any other; deltas stream in real time. Improvement Namespace creation is now O(1) — bulk agent provisioning from the SDK no longer blocks.
WebSocket streams + webhooks
Feature wss://remlabs.ai/v1/stream for live memory events. Subscribe per-namespace or globally; backpressure-aware.Feature HMAC-signed webhooks with 8 event types: memory.created, memory.updated, memory.deleted, dream.started, dream.completed, consolidation.ready, forecast.updated, namespace.shared. Improvement Retry policy on webhook delivery: exponential backoff with jitter, 24h retention.
Federation RBAC
Feature Shared spaces with per-user permissions — Owner, Admin, Developer, Viewer. Scoped at the memory, namespace, and channel levels. Improvement Namespace isolation audit logs. Every cross-namespace read is recorded, signed, and exportable as JSON Lines. Fix Dream Engine no longer leaks the cross-namespace embedding cache during consolidation passes.
MCP + A2A protocols
Feature Model Context Protocol server exposed at /mcp. Drop-in for Claude Desktop, Zed, and every MCP-native client. Feature Agent-to-Agent handshake at /a2a. Native agent discovery via Agent Card at /.well-known/agent-card.json. Improvement Both protocols share a unified auth layer — one API key, either transport.
Dream Engine GA
Breaking /v1/memory/ask response shape changed. The results array now nests under data.memories, and dream provenance is returned under data.dreams. See the migration guide ; the v3 shape is supported in parallel until 2026-07-04.Feature All 9 Dream Engine strategies production-ready: synthesize , pattern-extract , contradiction-detect , compress , associate , validate , evolve , forecast , reflect . Feature Per-strategy tuning knobs exposed on the POST /v1/dreams endpoint.
Self-host on Docker and Kubernetes
Feature docker-compose.yml published. Single-command self-host with SQLite backend, air-gap friendly.Feature Helm chart for Kubernetes deployments. Configurable storage class, ingress, and horizontal pod autoscaling. Improvement Cold-start boot time under 4 seconds on a 2-vCPU container.
80+ integrations shipped
Feature Automation : Zapier, n8n, Make — every memory action exposed as a trigger or action.Feature Coding : Cursor, Continue, Claude Code, Ollama — native MCP integration across every major agent IDE.Feature Frameworks : LangChain, LlamaIndex — first-class memory backend adapter in both.Feature Knowledge : Notion, Obsidian — two-way sync with wikilinks, tags, and folder structure preserved.
LongMemEval 94.6% benchmark
Feature Published full methodology at /benchmarks . 473/500 under byte-exact upstream GPT-4o judge — reproducible seed, dataset, and harness. Improvement FTS5 AND-first retrieval fix. Shaves 38ms off p95 recall latency without rebuilds. Improvement Leaderboard positions logged: AgentMemory 96.2, Chronos 95.6, Hindsight 94.6, REM 94.6.
Breaking-change policy
Breaking changes are announced 90 days in advance . The prior API version is supported for 180 days after deprecation . Subscribe to /changelog.rss for immediate notifications. API versioning follows semver: /v1/, /v2/. Deprecations are surfaced in Sunset and Deprecation response headers. See the versioning docs for the full policy.