Mneme — Semantic Memory for AI Agents

Mneme gives your AI agents persistent, queryable memory — with semantic vector search, a knowledge graph, and episodic tracking.

Named after the Greek goddess of memory, mother of the Muses.

Why Mneme?

AI agents forget everything between sessions. Mneme fixes that.

• Semantic search — find "coffee preference" when you ask about "drinks Doug likes" (no keyword matching required)
• Knowledge graph — entity triples with forward-chaining inference
• Episodes — session-level context so agents know what happened before
• Prompt-ready context — built-in context injector for any LLM prompt
• Web UI — inspect your memory at http://localhost:8765

ByteRover uses keyword search (BM25). Mneme uses semantic vectors. They're not the same thing.

Architecture

┌─────────────────────────────────────────────────────┐
│                    Your AI Agent                      │
│        (Claude Code, DeerFlow, Cursor, etc.)         │
└──────────────┬──────────────────────────────────────┘
               │ MCP tools / HTTP API
               ▼
┌──────────────────────────────────────────────────────┐
│                   Mneme Server                       │
│                  (FastAPI :8765)                     │
├──────────┬──────────┬──────────┬──────────────────┤
│  Facts   │  Graph   │ Episodes │  Vector (Chroma) │
│ SQLite   │ SQLite   │  SQLite  │   + Voyage AI    │
└──────────┴──────────┴──────────┴──────────────────┘

Quick Start

pip install mneme-ai

# Set your API key
export VOYAGE_API_KEY="pa-..."

# Start the server
mneme-server
# or: python -m mneme.server

# Visit http://localhost:8765 for the dashboard

MCP Server (recommended)

Connect Mneme to any MCP-compatible agent (Claude Code, Cursor, Cline, Windsurf):

# Install with MCP support
pip install "mneme-ai[mcp]"

# Start the MCP server
mneme-mcp

Then configure your agent to use the MCP server at localhost:8765.

API

Store a fact

curl -X POST http://localhost:8765/facts \
  -H "Content-Type: application/json" \
  -d '{"entity": "doug", "category": "preference", "content": "likes dark roast coffee", "importance": 7}'

Semantic search

curl "http://localhost:8765/vec/search?query=coffee+preferences&top_k=5"

Get full context for a prompt

curl -X POST http://localhost:8765/context/inject \
  -H "Content-Type: application/json" \
  -d '{"entity": "doug", "max_tokens": 2000}'

Knowledge graph

# What does the agent know about "doug"?
curl "http://localhost:8765/graph/know?who=doug"

Python API

from mneme import fact_store, vec_mem, entity_graph

# Store facts
fact_store.add("doug", "preference", "likes coffee", importance=7)

# Semantic search
results = vec_mem.search("coffee preferences", top_k=5)

# Knowledge graph
entity_graph.connect("doug", "likes", "coffee")

Memory Layers

Layer	What it does
Structured facts	SQLite, importance 1-10, decay over time
Raw messages	Immutable append-only log, checksums
Entity graph	Subject→predicate→object triples with inference
Episodes	Session grouping + event logging
Vector store	ChromaDB + Voyage AI (1024-dim embeddings)
Context injector	Token-budget-aware memory for prompts
Compaction	Distill → merge → prune → age-out

Configuration

Config file: ~/.config/mneme/config.json

{
  "env": {
    "VOYAGE_API_KEY": "pa-..."
  }
}

Or via environment variable: VOYAGE_API_KEY=pa-...

vs ByteRover

	ByteRover	Mneme
Search	BM25 keyword	Semantic vectors
Embeddings	None	Voyage AI (1024-dim)
Knowledge graph	No	Yes
Episodes	No	Yes
Agent integration	MCP	MCP + HTTP API
Auto-indexing	Manual	Auto on every write
Install	curl install.sh	pip install

License

MIT