A cognitive science-based memory framework for AI agents. Runs 100% local at $0 or with cloud APIs. Weibull forgetting, triple-path retrieval, 10-stage pipeline.
Other memory solutions store everything forever. Your agent drowns in noise. Context windows fill with stale facts. Retrieval degrades as data grows. The more your agent remembers, the worse it performs.
Mnemo models human memory: important memories consolidate, trivial ones fade, frequently recalled knowledge strengthens. Built on decades of cognitive science research, not naive vector search.
From raw conversation to durable, retrievable memory in milliseconds.
Every feature grounded in cognitive science and real-world agent workloads.
Stretched-exponential forgetting with tier-specific beta parameters. Memories fade naturally unless reinforced through recall.
Vector similarity, BM25 full-text, and knowledge graph traversal fused with Reciprocal Rank Fusion for robust recall.
Three-layer LLM detection pipeline. When facts conflict, old versions auto-expire and new truths consolidate.
Per-agent memory with configurable access rules. Each agent operates in its own namespace with controlled sharing.
Voyage rerank-2 precision ranking ensures the most relevant memories surface first, every time.
Spaced repetition, emotional salience scoring, and spreading activation. Memory that learns how to remember.
35 ablation tests prove every module earns its place. No cargo-cult features — each component has measurable, verified impact.
Core is free and fully functional. Cloud adds adaptive intelligence and zero-ops hosting.
$0 · MIT License
Free tier · Hosted API
Contact us
Self-host for free. Or use Mnemo Cloud for zero-ops.
Self-hosted, full framework
Hosted API, zero setup
Custom deployment and SLAs
Self-hosted Core: you bring your own API keys. Run 100% locally with Ollama for $0, or use cloud providers (~$5-45/mo). Ecosystem: Vercel AI SDK, REST Server, Python SDK.
# Install
npm install @mnemoai/core
# Pull Ollama model
ollama pull bge-m3
# Done! Use preset in your code:
const mnemo = await createMnemo({
preset: 'ollama',
dbPath: './memory-db'
});# Install
npm install @mnemoai/core
# Auto-detect from env var
const mnemo = await createMnemo({
dbPath: './memory-db'
});
# Or use a preset
const mnemo = await createMnemo({
preset: 'openai',
dbPath: './memory-db'
});import { createMnemo } from '@mnemoai/core';
// Initialize — auto-detect, preset, or full config
const mnemo = await createMnemo({
preset: 'openai', // or 'ollama', 'voyage', 'jina'
dbPath: './memory-db',
});
// Store a memory
await mnemo.store({
text: 'User prefers dark mode and concise responses',
category: 'preference',
});
// Retrieve — decay, dedup, and ranking applied automatically
const memories = await mnemo.recall('user preferences');
console.log(memories);
// [{ text: "User prefers dark mode...", score: 0.94, category: "preference" }]How Mnemo compares to other AI memory frameworks on architecture and features.
| Feature | Mnemo | Mem0 | Zep | Letta |
|---|---|---|---|---|
| Weibull Decay | Yes | No | No | No |
| Triple-Path Retrieval | Vec+BM25+Graph | Vector only | Vector+temporal | Vector only |
| Contradiction Detection | 3-layer LLM | Basic | No | No |
| Cross-encoder Rerank | Voyage/Jina/Ollama | No | No | No |
| 100% Local ($0) | Ollama preset | No | No | Partial |
| Open Source | MIT | Open core + SaaS | Paid | Freemium |
| Preset Config (1 line) | 4 presets | No | No | No |
| LOCOMO Benchmark | 85.2% (Cloud) / 46.4% (Core) | ~31.7%* | — | — |
| Ablation Validated | 35 tests | No | No | No |
| MCP Server | Built-in | No | No | No |
*Tested under identical conditions: same LOCOMO dataset, same GPT-4.1 judge, same scoring rubric. Open-source benchmark harness — reproduce it yourself.