MemBrain — The Cognitive Layer for AI¶
Memory, judgment, threat detection, and recall — the brain functions your LLMs are missing.
MemBrain is a self-hosted cognitive layer that sits between your applications and LLM providers. It gives you full visibility and control over every AI interaction — without changing a line of code.
What MemBrain does — the seven pillars¶
This is the canonical feature taxonomy. Every other doc in this site maps back to these pillars; if a feature isn't covered by one of them, it isn't shipped.
The community build is licensed under Apache 2.0. Features marked ★ Enterprise require a licensed add-on.
| Pillar | What it covers | Spec reference |
|---|---|---|
| Detection | 25+ PII patterns, fail-closed under scanner errors. Optional ML NER (BERT-based) with hybrid regex + NER mode — ★ Enterprise. | Configuration |
| Enforcement | Six policy modes (pass / log / alert / redact / block / confirm), tool-policy fnmatch globs, human-in-the-loop approval |
Configuration |
| Memory | Knowledge store on pgvector with semantic search; saved and reflected entries land pending and private, and require human approval and an explicit owner share before teammates see them, with attribution; PII rescan on injection | API Reference |
| Visibility | Audit log with a tamper-evident HMAC-SHA256 hash chain, encrypted PII mapping, GDPR export + right-to-erasure, Prometheus /metrics. Alert engine (webhook + Slack) — ★ Enterprise. |
API Reference |
| Routing | Multi-provider (Anthropic, OpenAI, Claude CLI, Ollama, LiteLLM 100+ models), tier / cost / privacy-based, fallback chains, exact + semantic response cache | Providers |
| Coverage | Three ingress modes — application proxy, transparent network proxy (TLS termination + SNI). MCP governance — ★ Enterprise. | Configuration |
| Trust | Multi-tenant isolation across cache / MCP registry / audit / knowledge, RBAC, peppered API-key hashing, atomic key rotation. OIDC SSO + SCIM — ★ Enterprise. | API Reference |
Quick Start¶
Point your AI client at MemBrain:
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8001/v1")
response = client.chat.completions.create(
model="gpt-4",
messages=[{"role": "user", "content": "Hello!"}]
)
See the Quickstart Guide for full setup instructions.
Documentation¶
| Section | Description |
|---|---|
| Quickstart | Get up and running in 5 minutes |
| Configuration | All settings and environment variables |
| API Reference | Complete endpoint documentation |
| Providers | Supported AI providers and routing |
| Python SDK | Drop-in OpenAI SDK replacement |
| Team knowledge sharing | Two-gate model — saved and reflected entries land pending and private; a human approves, then the owner explicitly shares with the team; shared entries carry owner attribution |