Most AI tools forget everything between sessions. MitrAI Memory Engine gives your AI a persistent, searchable knowledge base that automatically improves its own retrieval quality over time.
Upload documents, ask questions, and watch the system learn your domain — powered by DSPy optimization that compiles per-user query strategies from your actual data.
Why it's different
Self-Improving Retrieval
DSPy autotuner generates training data from your uploads and compiles optimized query strategies per user. Your knowledge base literally gets smarter.
Any LLM Provider
Ollama, OpenAI, Anthropic, Cohere, and 100+ more via LiteLLM. One config change to swap providers. No code changes, no lock-in.
Local-First Privacy
Run entirely on your machine with Ollama. Your documents, embeddings, and optimized models never leave your hardware.
Production Architecture
Qdrant vector store, per-user collection isolation, Bearer token auth, health probes, and graceful degradation built in.
Three ways to connect
15 endpoints for web apps, mobile, and custom integrations. Full OpenAPI docs included.
curl localhost:8100/api/query
9 tools for Claude Desktop, Cursor, OpenClaw, and any MCP-compatible AI agent.
memory-mcp
Interactive setup wizard, health checks, ingest, query — all from your terminal.
memory-engine init
Works with your agent stack
One memory backend. Every agent framework. Pick your integration path.
Claude Desktop
Add memory to Claude as an MCP tool. Ingest docs, query knowledge, manage collections — Claude calls the tools automatically.
{
"mcpServers": {
"memory-engine": {
"command": "memory-mcp",
"env": { "MEMORY_API_URL": "http://localhost:8100" }
}
}
}Cursor
Give Cursor persistent project memory. Index codebases, design docs, and ADRs — get contextual answers in the editor.
{
"mcpServers": {
"memory-engine": {
"command": "memory-mcp",
"args": ["--user", "cursor-project"]
}
}
}OpenClaw
Native OpenClaw plugin with 8 tools. Install from the marketplace or point to a local build. Full read/write memory access.
{
"name": "mitrai-memory",
"tools": ["ingest", "query", "search",
"list_collections", "delete_document",
"get_stats", "health_check", "optimize"]
}Agent Frameworks
LangChain
Use as a custom retriever. Two HTTP calls: /api/ingest-batch to store, /api/query to retrieve. Fits into any LangChain chain.
from langchain.retrievers import BaseRetriever
import httpx
class MemoryRetriever(BaseRetriever):
def _get_relevant_documents(self, query):
r = httpx.post("http://localhost:8100/api/query",
json={"query": query, "user_id": "my-agent"})
return [Document(page_content=c["content"])
for c in r.json()["results"]]CrewAI
Give CrewAI agents long-term memory across tasks. Each crew member gets its own user namespace with isolated collections.
from crewai import Agent
from crewai_tools import tool
@tool("search_knowledge")
def search(query: str) -> str:
"""Search the team knowledge base."""
r = httpx.post("http://localhost:8100/api/query",
json={"query": query, "user_id": "crew-research"})
return r.json()["results"][0]["content"]AutoGen / AG2
Register memory as a function tool. AutoGen agents call it automatically when they need context from uploaded documents.
assistant.register_for_llm(
name="query_memory",
description="Search the knowledge base"
)(lambda query, user_id="autogen-team":
httpx.post("http://localhost:8100/api/query",
json={"query": query, "user_id": user_id}
).json()["results"]
)Any Agent / Custom
15 REST endpoints. If your framework can make HTTP calls, it can use the Memory Engine. Bearer token auth, JSON in/out, OpenAPI spec.
curl -X POST http://localhost:8100/api/query \
-H "Authorization: Bearer $API_KEY" \
-H "Content-Type: application/json" \
-d '{"query": "quarterly revenue",
"user_id": "my-agent",
"top_k": 5}'Built for real workflows
Agent Memory
Give any AI agent persistent memory across sessions. Ingest context once, recall it forever. Each agent gets its own namespace.
Team Knowledge Base
Ingest company docs, wikis, and SOPs. Every team member's AI assistant shares the same searchable knowledge.
Code Context
Index repos, ADRs, and design docs. Get contextual answers about your codebase inside Cursor or any MCP-enabled editor.
Research & Learning
Drop papers and notes. The engine builds connections across your reading and surfaces relevant context as you ask questions.
Provider presets
Start with a preset, customize with environment variables.
Try the Memory Engine with Mitrai App
Mitrai is a private, local AI chat app with the Memory Engine built in. Upload documents, ask questions, and see self-improving retrieval in action — no integration required.
💬 Chat with Documents
Drop files into the conversation. Memory Engine indexes them and returns cited answers.
🧠 Self-Improving Answers
The more you use it, the better retrieval gets. DSPy optimization runs in the background.
🔒 Completely Local
Ollama models, local vector store, no data leaves your machine. One-click install.
You can use the cloud version to see how the engine works. However, if you prefer, we have a downloadable Private Chat version.
Feature demos (YouTube series)
Short, focused videos — coming soon. We'll add them here feature-by-feature.