You explained your entire architecture to Claude. Two hours of context, decisions, tradeoffs. Next day, new conversation — it knows nothing. So you explain it again. Then you switch to Cursor for implementation. Zero context. You open ChatGPT to brainstorm a different angle. Blank slate.
Persistent memory graph for AI agents. Facts, decisions, entities, and relationships that survive across sessions, tools, and providers. MCP server — works with Claude, Cursor, ChatGPT, and any MCP client.
ดาว
44
เครื่องมือ
—
อัปไทม์
—
Every AI agent you use is brilliant but amnesiac. And none of them talk to each other.
MIE fixes this. One knowledge graph. Every agent reads from it. Every agent writes to it. Your decisions, your context, your rules — always available, everywhere.
How It Works
You: "We chose PostgreSQL over DynamoDB because we need ACID
transactions for the payments module. Alternative was Aurora
but too expensive at current stage."
Claude stores this via MIE
↓
┌─────────────────────────────────────────────┐
│ MIE Knowledge Graph │
│ │
│ Decision: PostgreSQL over DynamoDB │
│ Rationale: ACID transactions for payments │
│ Alternatives: [DynamoDB, Aurora] │
│ Entities: payments-module, PostgreSQL │
│ Status: active │
└─────────────────────────────────────────────┘
↓
Next week, in Cursor, different project:
"What database did we choose and why?"
→ Cursor queries MIE, gets full context instantly
No copy-pasting. No "as I mentioned before." No starting from zero.
Why Not Just Use Claude's Memory / ChatGPT's Memory?
These connect through typed relationships — a decision references entities, relates to topics, and may be triggered by events. When an agent queries "what do you know about our security decisions?", MIE traverses the graph and returns structured context, not keyword matches.
MCP Tools
MIE exposes 12 tools through the Model Context Protocol:
Tool
What it does
mie_analyze
Surfaces related context before storing — the agent decides what's worth remembering
mie_store
Writes facts, decisions, entities, events, and relationships to the graph
mie_bulk_store
Batch store up to 50 nodes with cross-references — ideal for importing knowledge from files or git history
mie_get
Retrieve a single memory node by ID with full details
mie_query
Semantic search, exact lookup, or graph traversal across all node types
mie_list
List and filter nodes with pagination
mie_update
Invalidate outdated facts, update statuses — with full history preserved
mie_delete
Remove nodes with cascade (embedding + edges) or remove individual relationships
mie_conflicts
Detect contradictions in stored knowledge
mie_export
Export the full graph as JSON or Datalog
mie_repair
Rebuild HNSW indexes and clean orphaned embeddings
mie_status
Graph health, node counts, usage metrics
Zero Server-Side Inference
Unlike other memory solutions that run an LLM on the server to classify what to store, MIE uses an agent-as-evaluator pattern. The server provides context; your agent (which is already running an LLM) decides what matters. This means zero additional inference cost — your memory layer doesn't burn tokens.
This philosophy extends to importing: when you ask your agent to "import knowledge from this repo", the agent reads your files, ADRs, or git history directly and uses mie_bulk_store to persist what it extracts. MIE stays as a pure storage engine — the connected agent IS the LLM.
ChatGPT via custom GPT Actions pointing to MIE Cloud (coming soon).
Multiple MCP clients (Claude, Cursor, etc.) can run simultaneously — the daemon holds the exclusive database lock and multiplexes access. The daemon starts automatically on first use or can be managed manually via mie daemon.
Memory Lifecycle
Store Query Evolve
───── ───── ──────
Your agent learns → Next session, any → Facts change.
something new. agent queries MIE Old ones get
It stores a fact, for context before invalidated, not
a decision, or an responding. Full deleted. The graph
entity — with graph of related keeps history of
confidence scores knowledge returns what was known
and relationships. in milliseconds. and when.
Configuration
# .mie/config.yaml
version: "1"
storage:
engine: rocksdb # rocksdb, sqlite, or mem
embedding:
enabled: true
provider: ollama # ollama, openai, or nomic
model: nomic-embed-text
mcp:
# Option A: only expose these tools (whitelist)
include_tools:
- mie_analyze
- mie_store
- mie_query
# Option B: expose all except these (blacklist)
# exclude_tools:
# - mie_export
# - mie_repair
All settings can be overridden with environment variables. Embeddings are optional — MIE works without them (exact search only).
The mcp section is optional. When omitted, all 12 tools are exposed. Use include_tools to whitelist specific tools, or exclude_tools to hide a few. If both are set, include_tools takes precedence. Filtered tools are also rejected on tools/call.
CLI
mie init # Create config with defaults
mie init --interview # Interactive project bootstrapping
mie --mcp # Start as MCP server (auto-starts daemon)
mie daemon start # Start daemon in background
mie daemon start --foreground # Start daemon in foreground (for debugging)
mie daemon stop # Stop running daemon
mie daemon status # Check if daemon is running
mie status # Show graph statistics
mie export # Export memory graph
mie import -i backup.json # Import from JSON or Datalog
mie reset --yes # Delete all data
mie query "<cozoscript>" # Raw Datalog query (debug)
Prerequisites
Go 1.24+ (building from source)
Ollama (optional, for semantic search) — ollama pull nomic-embed-text
MIE works without Ollama. You get exact-match search and graph traversal. Add Ollama for semantic search ("find things related to deployment" instead of exact keywords).
CIE gives your agent deep understanding of your codebase. MIE gives it memory of everything else — decisions, architecture, people, events. Together, your agent knows your code and remembers why it's built that way.
Roadmap
Import from ADRs, markdown, and git history (agent-driven self-import)
Git post-commit hook — auto-capture decisions from commits
Browser extension — auto-capture knowledge from claude.ai, chatgpt.com, gemini
MIE Cloud — sync across devices, team shared memory
ChatGPT integration via custom GPT Actions
Web UI for exploring and managing your knowledge graph
Persistent memory graph for AI agents. Facts, decisions, entities, and relationships that survive across sessions, tools, and providers. MCP server — works with Claude, Cursor, ChatGPT, and any MCP cl
Persistent memory graph for AI agents. Facts, decisions, entities, and relationships that survive across sessions, tools, and providers. MCP server — works with Claude, Cursor, ChatGPT, and any MCP client.
ตรวจสอบล่าสุดเมื่อ 2026-10-28 Persistent memory graph for AI agents. Facts, decisions, entities, and relationships that survive across sessions, tools, and providers. MCP server — works with Claude, Cursor, ChatGPT, and any MCP client.