Mnemo

Long-term memory for AI agents.
Store, recall, and forget — just like humans do.

Quick Start · Docs · Why Mnemo · Self-hosted vs Cloud · Website

---

Quick Start

npm install @mnemoai/core

import { createMnemo } from '@mnemoai/core';

const mnemo = await createMnemo({ dbPath: './memory-db' });

// Store
await mnemo.store({ text: 'User prefers dark mode and minimal UI' });

// Recall — vector search + BM25 + rerank + decay scoring
const results = await mnemo.recall('What does the user like?');
// → [{ text: "User prefers dark mode and minimal UI", score: 0.92 }]

// Old memories fade automatically. Important ones stick around.

Auto-detects OPENAI_API_KEY from env. Or use a preset:

// 100% local, $0 API cost
const mnemo = await createMnemo({ preset: 'ollama', dbPath: './memory-db' });

Available presets: openai · ollama · voyage · jina — configuration guide

Python

pip install mnemo-memory
npx @mnemoai/server   # start the REST API

from mnemo import MnemoClient

client = MnemoClient()
client.store("User prefers dark mode", category="preference")
results = client.recall("UI preferences")

100% Local with Ollama ($0)

ollama pull bge-m3               # embedding
ollama pull qwen3:8b             # smart extraction LLM
ollama pull bge-reranker-v2-m3   # cross-encoder rerank

const mnemo = await createMnemo({ preset: 'ollama', dbPath: './memory-db' });

Full Core functionality — embedding, extraction, rerank — all running locally.

Docker (full stack with Neo4j + Dashboard)

git clone https://github.com/Methux/mnemo.git
cd mnemo
cp .env.example .env     # add your API keys
docker compose up -d     # starts Neo4j + Graphiti + Dashboard

---

Why Mnemo?

Most AI memory systems are glorified vector databases — they store everything and retrieve by similarity. That breaks at scale: your agent drowns in stale, contradictory, and irrelevant memories.

Mnemo is different. It models memory the way cognitive science says humans actually remember:

• Old memories fade. A Weibull decay model naturally deprioritizes stale information — no manual cleanup needed.
• Important memories consolidate. Frequently accessed, high-importance memories promote to a "core" tier with slower decay.
• Contradictions resolve automatically. When a user says "I moved to Tokyo" after previously saying "I live in NYC", Mnemo detects the contradiction and expires the old fact.
• Noise gets filtered. Debug logs, API errors, meta-questions — automatically excluded from long-term storage.

The result: your agent's memory stays sharp at 100 memories or 10,000.

How it compares

	Mnemo	Mem0	Zep	LangMem
Local-first (no SaaS lock-in)	Yes	Yes	CE deprecated	Partial
Forgetting model	Weibull decay	None	Time window	None
Contradiction detection	3-layer LLM	Graph layer	Temporal versioning	None
Multi-backend (LanceDB/Qdrant/Chroma/PGVector)	Yes	Qdrant	Postgres	Varies
Provider agnostic (BYO embedding/LLM)	Yes	Limited	No	LangChain only
Fully offline ($0 with Ollama)	Yes	Partial	No	No
Cross-encoder rerank	Yes	Yes	Yes	No

---

Architecture

  User message
       │
       ▼
  ┌─── Store ───────────────────────────────────────┐
  │  Embed → Noise filter → Dedup → Contradiction   │
  │  detection → LanceDB (vector + BM25 index)      │
  └──────────────────────────────────────────────────┘
       │
       ▼
  ┌─── Recall ──────────────────────────────────────┐
  │  Vector search + BM25 → RRF fusion → Rerank     │
  │  → Decay scoring → MMR diversity → Top-K         │
  └──────────────────────────────────────────────────┘
       │
       ▼
  ┌─── Lifecycle ───────────────────────────────────┐
  │  Working → Core (consolidate)                    │
  │  Working → Peripheral → Archive (fade out)       │
  │  Driven by composite score, no manual tuning     │
  └──────────────────────────────────────────────────┘

Every parameter adapts to your store size. No magic numbers to tune.

---

Feature Highlights

Capability	Core (Free)	Cloud
Vector + BM25 + Knowledge Graph	Yes	Yes
Weibull forgetting model	Yes	Yes
Memory tiers (Core/Working/Peripheral)	Yes	Yes
Cross-encoder rerank	Yes	Yes
Contradiction detection	Yes	Yes
Multi-backend (LanceDB, Qdrant, Chroma, PGVector)	Yes	Yes
Scope isolation (multi-agent)	Yes	Yes
$0 local deployment (Ollama)	Yes	Yes
Adaptive retrieval (pool/score/frequency)	—	Yes
Extraction-time context injection	—	Yes
Session deduplication	—	Yes

---

Self-hosted vs Cloud

Self-hosted (Core) — the full framework, MIT licensed, no restrictions. npm install @mnemoai/core and run it yourself. You bring your own embedding/LLM keys.

Mnemo Cloud — hosted API, zero setup. Adaptive retrieval, intelligent extraction, and contradiction detection built in. No keys to manage, no infrastructure to run.

npm install @mnemoai/client

import { createCloudMnemo } from "@mnemoai/client";

const mnemo = createCloudMnemo({ apiKey: "mn_your_key" });
await mnemo.store({ text: "User prefers dark mode" });
const memories = await mnemo.recall("UI preferences");

Plan	Price	What you get
Core	Free forever	Full framework, self-hosted, MIT
Cloud Starter	$29/month	10K memories, 1K stores/day, 50K recalls/day
Cloud Pro	$99/month	100K memories, 10K stores/day, unlimited recalls
Enterprise	Contact us	Custom limits, dedicated support, SLA

Sign Up for Mnemo Cloud →

API Configuration Guide

Mnemo is a framework — you bring your own models. Choose a setup that fits your budget:

Setup	Embedding	LLM Extraction	Rerank	Est. API Cost
Local	Ollama bge-m3	Ollama qwen3:8b	Ollama bge-reranker	$0/mo
Hybrid	OpenAI text-embedding-3-small	GPT-4.1-mini	Jina reranker	~$5/mo
Cloud	Voyage voyage-4	GPT-4.1	Voyage rerank-2	~$45/mo

These are your own API costs, not Mnemo subscription fees.

---

Packages

Package	Platform	Install
@mnemoai/core	npm	npm install @mnemoai/core
@mnemoai/client	npm	npm install @mnemoai/client
Mnemo Cloud	REST API	Sign up at api.m-nemo.ai/signup
@mnemoai/server	npm	npx @mnemoai/server
@mnemoai/vercel-ai	npm	npm install @mnemoai/vercel-ai
mnemo-memory	PyPI	pip install mnemo-memory

---

Cognitive Science

Mnemo's design maps directly to established memory research:

Human Memory	Mnemo
Ebbinghaus forgetting curve	Weibull decay model
Core vs peripheral memory	Tier system with differential decay rates
Interference / false memories	Deduplication + noise filtering
Metamemory	mnemo-doctor + Web Dashboard

Read more: Architecture → · Retrieval Pipeline → · Ablation Tests →

---

Documentation

Full docs at docs.m-nemo.ai

• Quick Start
• Local Setup ($0 Ollama)
• Configuration Reference
• Storage Backends
• Retrieval Pipeline
• API Reference

---

Tools

Tool	Description	Run
mnemo init	Interactive config wizard	npm run init
mnemo-doctor	One-command health check	npm run doctor
validate-config	Config validation gate	npm run validate
Dashboard	Web UI for browsing, debugging, monitoring	http://localhost:18800

---

Contributing

We welcome contributions to Mnemo Core (MIT-licensed files). See CONTRIBUTING.md.

Areas where we'd love help:

• Benchmark evaluation (LOCOMO, MemBench)
• New storage adapters and embedding providers
• Retrieval pipeline optimizations
• Documentation and examples

---

License

Dual-license model:

• MIT — Core framework (SPDX-License-Identifier: MIT)
• Commercial — Cloud features and advanced strategies

See LICENSE for details.

---

_{Built with cognitive science, not hype.}

---

_{**Trademarks:** LanceDB is a trademark of LanceDB, Inc. Neo4j is a trademark of Neo4j, Inc. Qdrant is a trademark of Qdrant Solutions GmbH. Mnemo is not affiliated with, endorsed by, or sponsored by any of these organizations. Storage backends are used under their respective open-source licenses.}