# Memra

> Memra is a persistent memory API for AI agents and LLM applications. EU-native (hosted in Helsinki, Finland), privacy-first, deterministic sub-100ms recall with no LLM in the hot path. REST API + MCP server + Python/TypeScript/PHP SDKs. Store facts, decisions, patterns and preferences; recall them by hybrid semantic + lexical search with staleness and trust signals on every result.

Memra is built for correctness: supersede chains retire outdated facts (and their derived children), deletes cascade provably across index, files and caches, every write returns a read-your-writes revision token, and write-time contradiction detection warns an agent when a new fact clashes with stored knowledge. PII masking (Presidio-based, 7 EU language models) is available per project. No LLM sits in the recall path — results are reproducible and fast (p50 ≈ 40ms).

Sovereign mode (Team add-on, €99/mo; included with Enterprise): per-project zero non-EU subprocessors. Sovereign projects embed with a local multilingual model (multilingual-e5-small, ~100 languages) on Memra's own EU servers — no OpenAI, no Anthropic — so memory content never leaves the server. Set `tool_egress_policy: "restricted"` at project creation (immutable afterwards).

## Getting started

- [Install & quickstart](https://usememra.com/install): create an account, get an API key (`memra_live_...`), store your first memory in 2 minutes
- [OpenAPI specification](https://usememra.com/docs/openapi.json): complete machine-readable REST API reference
- [Python SDK](https://usememra.com/docs/sdks/python): `pip install memra-sdk`
- [TypeScript SDK](https://usememra.com/docs/sdks/typescript): `npm install @usememra/sdk`
- [PHP / Laravel SDK](https://usememra.com/docs/sdks/php): `composer require memra/sdk`

## API essentials

- Base URL: `https://usememra.com/api/v1` — Bearer auth with `memra_live_` keys
- Core endpoints: `POST /memories` (returns `revision` + `conflicts[]`), `POST /memories/recall` (hybrid semantic+lexical, accepts `wait_for_revision`, `max_tokens`, `not_tags`, `since`, `until`, `used_ids`), `POST /memories/{id}/supersede`, `GET /entities`, `POST /memories/feedback`
- MCP server (Streamable HTTP): `https://usememra.com/mcp` — tools: memra_remember, memra_recall, memra_get, memra_list, memra_supersede, memra_history, memra_bootstrap
- Every recall result carries: `score`, `staleness_score` (0–100), `staleness_status`, `last_confirmed`, `trust_state` — the agent decides what to trust
- Read-only scoped API keys let a lead agent hand subagents recall-only access

## Product

- [Features](https://usememra.com/features): memory types, supersede & audit trails, PII masking, multi-tenant namespaces, entity graph
- [Pricing](https://usememra.com/pricing): flat monthly tiers, no per-token metering
- [Security](https://usememra.com/security): key handling, tenant isolation, deletion guarantees
- [Data & privacy](https://usememra.com/data-privacy): EU data residency (Hetzner, Helsinki), privacy-first processing
- [Benchmark](https://usememra.com/benchmark): honest, reproducible recall benchmarks
- [Changelog](https://usememra.com/changelog): release history

## Integrations

- [Hermes Agent memory provider](https://usememra.com/hermes-memory): hermes-memra plugin, fully-local mode supported
- [memra-local](https://pypi.org/project/memra-local/): `pip install memra-local` — offline local memory server (SQLite + ONNX embeddings) with the same MCP tools
- [pi coding agent extension](https://www.npmjs.com/package/@usememra/pi-memory): hybrid cloud + local memory

## Optional

- [About](https://usememra.com/about): why Memra exists, who builds it
- [Terms](https://usememra.com/terms)
- [llms-full.txt](https://usememra.com/llms-full.txt): full API reference for LLM consumption