Persistent
Durable semantic and procedural notes, backed by MariaDB metadata and Qdrant vector retrieval.
Early access · open to join
Blue Note Memory
MCP-native memory infrastructure for AI agents: durable semantic memory, episodic logs, scratch state, procedural/todo memory, graph lookup, scoped ownership, and safe secret handling. Connect an agent in minutes.
Founding pricing from $19/mo · self-serve · cancel anytime · SaaS, hybrid & on-prem.
MCP-nativePersistent, scratch, episodic, feedback, proposal, and graph memory over one HTTP tool adapter — on every plan.
Scoped ownersUser, team, agent, project, client, tenant, and admin memory with audited sensitive reads.
Secret-safeRaw secrets are blocked at write time; memory stores vault_key, secret_ref, or env references only.
Yours to keepRun it hosted, hybrid, or fully on-prem. Export or delete your data anytime.
Agent memory, not prompt sprawl
The infrastructure remembers.
Agents should not carry the burden of long-term recall inside one context window. Blue Note Memory assembles the right scoped memory when a session starts, logs the important episode afterward, and lets humans approve durable changes.
Launch surface
Memory tiers built for real agent operations.
Episodic
Append-only event logs for sessions, handoffs, tool runs, and post-turn voice-agent summaries.
Scratch
Short-lived working memory for queue handoffs, active investigations, and task-local context.
Todos
Structured procedural memory for open, blocked, done, and verified tasks with owner scopes.
Graph
Apache AGE graph facts for relationships that keyword and vector search do not naturally surface.
Learning loop
Feedback and proposals keep agent learning human-reviewed instead of silently rewriting memory.
Beyond memory
Routing, skills, and tools — governed, and on every plan.
Harness routing
Per-phase model routing (plan / write / test) with a per-tenant LiteLLM gateway key — point each coding phase at the right model without hardcoding providers.
Governed skills
Versioned instruction playbooks served from memory, identical in every host; prompt-kind skills surface as native MCP prompts. Script-bearing skills wait for your approval before they run.
Custom tools
Author read-only custom tools — a reusable prompt template or a preset memory search — callable by slug over MCP. No egress, no code execution.
Data plane
MariaDB, Qdrant, Apache AGE, and policy gates.
The hosted service keeps the logical stores self-contained on the same operating footprint while preserving clear boundaries between auth records, vector indexes, graph facts, and billing mirrors.
