Skip to main content
AI assistants are stateless by default. Every session starts blank — no memory of past decisions, no awareness of ongoing projects, no recollection of the people and rules that matter to you. You repeat yourself constantly. Your AI never learns. Memory Crystal fixes this. It’s a persistent cognitive memory layer that hooks into your AI client, captures every conversation, extracts durable facts and decisions, and automatically injects relevant context before each response. Your AI finally remembers.

​
How it works

When you send a message, Memory Crystal’s Context Engine runs before the AI responds. It searches both short-term and long-term memory, ranks the most relevant results, and injects them into the model’s context window — all automatically. After the AI responds, Memory Crystal captures the conversation and uses a secondary model to extract durable memories: decisions made, lessons learned, facts established, goals set. Those memories are stored permanently and become available in every future session.

​
Two memory layers

Memory Crystal maintains two distinct layers that work together on every query:
LayerWhat it storesRetention
Short-term memory (STM)Raw messages, verbatimRolling window (7–90 days depending on plan)
Long-term memory (LTM)Extracted facts, decisions, lessons, people, and rulesPermanent, vector-indexed
STM gives your AI perfect short-term recall of recent conversations. LTM gives it permanent knowledge that persists across sessions indefinitely. Both are searched on every turn.

​
Three deployment modes

Choose how you want to run Memory Crystal:

Cloud

Hosted at memorycrystal.ai. Zero configuration — install and go. Managed infrastructure, automatic updates.

Self-hosted

Deploy your own Convex backend. Full data sovereignty — your memories never leave your infrastructure.

Local

SQLite only, no cloud. Your data stays on your machine. Context engine only — no cloud sync.

​
Who Memory Crystal is for

  • Individual developers who want their AI coding assistant to remember past decisions, learned patterns, and project context across sessions
  • Teams who need AI agents with shared, scoped memory across multiple collaborators and workstreams
  • Anyone using MCP-compatible AI clients — Claude Desktop, Cursor, Windsurf, Claude Code, and others

​
Feature highlights

  • Automatic capture — every conversation is captured without any manual action
  • Intelligent extraction — a secondary model distills durable memories from raw conversation
  • Knowledge graph — related memories are connected so your AI understands relationships, not just isolated facts
  • 24 memory tools — search, recall, checkpoint, trace, and manage memories directly from your AI client
  • Knowledge bases — import stable reference material like runbooks, policies, and docs alongside conversational memory
  • Multi-tenant isolation — channel scoping keeps memories private to the right workspace, client, or agent
  • Prompt injection protection — recalled memories are sanitized before injection to prevent adversarial content
Memory Crystal ships as an OpenClaw plugin, an MCP server for any compatible host, and a web dashboard. You can use it with Claude Desktop, Cursor, Windsurf, Claude Code, or any MCP-compatible client.
Ready to get started? Follow the Quick Start to have Memory Crystal running in minutes.