what's new in memra

+ memra-local published to pypi — pip install memra-local (busl-1.1, flips to apache-2.0 on 2030-04-17)

+ @usememra/pi-extension published to npm — persistent memory for the pi coding agent

+ install page expanded: droid (factory.ai), hermes agent, openclaw added alongside claude code, cursor, and codex

+ tenant isolation: cross-account regression test suite added — one account can never read another's memories

+ mcp schema hardening: outputSchema roots normalised to type: object per spec 2025-06-18 (fixes client compatibility)

+ new primary mcp verbs: memra_remember and memra_recall

+ visible tool surface narrowed from 12 to 7; legacy names still callable until 2026-06-30

+ structured outputs with structuredContent + text fallback (mcp spec 2025-06-18)

+ canonical memory uris: memra://project/{pid}/namespace/{tid}/memory/{mid}

+ every retrieved memory carries trust_state, content_origin, untrusted_content flags for prompt-injection safety

+ per-tool-family rate limits (remember / recall / bootstrap / history / read)

+ audit trail persisted for every mcp tool call (params hashed, 90-day retention)

+ scaffolded tool_egress_policy per project (enforcement lands in v4.4)

~ most mcp clients cache tool lists at connect — restart your client to see the new 7-tool surface

+ memra local: self-hosted memory server with mcp and http interfaces for offline and on-device use

+ new install page with local and cloud setup side by side

+ new memra_supersede, memra_history, and memra_promote mcp tools for memory evolution

+ python, typescript, and php sdk quickstart guides

+ estimated token usage surfaced on recall and search responses

+ reranking on by default for higher recall quality

+ dashboard: connect memra local onboarding card and per-project api examples

+ automatic content compression for large memories to reduce token footprint

+ 90-day dormant memory retention for cancelled accounts

+ mcp transport reliability improvements and request tracing

+ additional security headers and content security policy hardening

+ new memra_get mcp tool for full memory retrieval by id

+ sdk install commands on dashboard for all 3 languages

+ auth pages redesigned with brand styling and dark mode

+ dark mode polish across dashboard surfaces

+ usage chart now tracks actual memory count over time

+ playground page layout improvements

+ revoked api keys hidden from dashboard and key count

+ svg dual-circle background textures for visual depth

+ hero section glow effect with gradient overlays

+ interactive code snippet tabs (php, python, typescript, curl)

+ social proof stat cards with animated counters

+ dark mode improvements across all marketing pages

+ design system with dark/light mode and 24 reusable components

+ public marketing pages: pricing, features, about, security, privacy, changelog

+ data privacy pages with sub-processor transparency

+ privacy shield: automatic pii detection and deterministic masking

+ native vector search with optimized retrieval

+ entity graph relationships for connected memory traversal

+ hybrid retrieval scoring for higher recall quality

+ memory classification and automatic type detection

+ importance scoring with configurable thresholds

+ staleness detection and health scoring

+ intelligent recall with context-aware ranking

+ python sdk published on pypi with sync and async support

+ typescript sdk with native fetch, edge-compatible

+ php sdk finalized for packagist with privacy helpers

+ recall performance optimization (sub-200ms latency)

+ locomo benchmark integration for recall quality tracking

+ core memory api: crud, batch (up to 100), 8 memory types

+ semantic recall with hybrid retrieval

+ multi-tenant isolation with full deletion cascade

+ subscription billing with 4 tiers and usage metering

+ customer dashboard with api key management

+ laravel sdk for php developers

+ mcp server with 8 tools for ai agent integration

+ openapi 3.0 spec with interactive api docs