A semantic memory database your agents can read, write, and reason over — across sessions, models, and teams. Model-agnostic. Enterprise-secure. Native MCP and REST API.
› remember why we picked Stripe over Adyen memory_search → 3 results · bank: payments-platform [decision · importance 9] Chose Stripe for global card coverage + Issuing API; Adyen contract required $200k/yr minimum and lacked sub-account model we needed for multi-tenant payouts. Reviewed 2026-02-11 with finance + legal. [constraint] Never log raw PAN or CVV — PCI scope must stay frozen. › remember this: rate-limited Stripe webhook to 50/s, was 502'ing on bursts memory_remember → saved · type: pattern · talked-like-ai ✓
Model-agnostic — works with whatever you use
Features
A purpose-built memory layer for agentic systems — semantic search, multi-tenant security, and a credentials gateway so your agents never hold the keys.
Drop in raw notes, Slack threads, or messy bullet points. Memory System auto-rewrites them into AI-optimized format — structured, retrievable, and tuned for LLM context windows. Your humans write like humans; your agents read like agents.
Claude, GPT, Gemini, Llama, Mistral, your local Ollama — same memory bank, every agent. Switch models without losing context.
AES-256-GCM at rest, JWT + MFA, row-level multi-tenant isolation, full audit logging, rate limiting, and a SOC 2 readiness path.
First-class Model Context Protocol server for Claude Code and a clean REST API for everything else. Drop in via one config block.
Agents call Cloudflare, AWS, and other APIs through the gateway — they never see the secret. Fine-grained scopes, elevation workflows, full audit trail.
Hierarchical departments, cross-team bank sharing with granular CRUD permissions, and time-boxed access for contractors.
See which memories actually get retrieved, which drift stale, and where your agents are flying blind.
How it works
Add a single MCP block to Claude Code, or hit the REST API from any agent in any language.
Your agent calls memory_remember and memory_search. We auto-summarize, embed, and structure into AI-readable form.
Departments, bank sharing, fine-grained scopes, elevation requests — Memory System grows with your org.
{
"mcpServers": {
"memory-system": {
"command": "npx",
"args": ["memory-system-mcp"],
"env": {
"MEMORY_API_KEY": "ms_xxxxx"
}
}
}
}
Security & compliance
Memory is sensitive. We treat it like the production data store it is.
Every memory body encrypted with envelope keys. TLS 1.3 in flight.
Short-lived 15-min access tokens, refresh rotation, optional MFA enforced org-wide.
Postgres RLS, per-org keys, departments, scoped API keys.
Append-only audit log with export, plus elevation approval workflow.
Hierarchical namespaces, versioning, agents act through the gateway — never holding keys.
Controls mapped from day one. HIPAA/BAA on Enterprise.
Pin your data where your customers expect it.
Docker images for VPC, air-gapped, or fully on-prem deployments.
Pricing
All plans include MCP, REST API, encryption at rest, and "talks like AI" auto-formatting.
For one developer running agents day-to-day.
Start free trialEverything a team needs to run shared agentic memory.
Start free trialCompliance, on-prem, dedicated infra.
Talk to salesOverage on Pro: $0.50 per additional GB · $1.50 per additional 1M embedding tokens. No charge for memory reads.
Whatever you save — a one-line note, a pasted Slack thread, a half-finished thought — gets auto-rewritten into AI-optimized form: typed (decision/pattern/constraint), summarized, tagged, and embedded. So when an agent retrieves it later, it lands in context windows ready to use, not as raw mush.
All of them. Native MCP for Claude Code, plus a clean REST API for OpenAI, Gemini, Llama, Cursor, Cline, LangChain, custom agents — anything that can make an HTTP call.
A vector DB stores embeddings. Memory System is the full memory layer above it: typed memories, summarization, departments, audit, encryption, MCP tooling, a service gateway, and the AI-formatting pipeline. We use pgvector underneath — you don't have to think about it.
Both. The Service Gateway lets agents execute scoped API actions (Cloudflare DNS, AWS, more coming) without ever holding the credentials themselves. Sensitive actions go through an elevation/approval flow.
AWS us-east-1 by default. Pro and Enterprise customers can pin to EU. Enterprise can deploy on-prem or in your VPC.
No. Ever. Your memories are encrypted at rest with per-org keys. We don't train on customer data and we don't share it with model providers beyond the embedding/summarization calls you've explicitly enabled.
Spin up a memory bank in minutes. 14-day free trial on every plan, no credit card required.