AI 开发中最消磨精力的，莫过于这种被迫的“上下文重置”——只要切换环境，之前做过什么、卡在哪个环节、当时的思考逻辑是什么，全得重新交代清楚。

针对这个痛点，有人推出了 Memanto，一个专为 AI 打造的工作记忆库，目前已支持 Claude Code、Codex、Cursor、LangGraph、CrewAI 等超过16种开发环境。

它并不是简单地把聊天记录存下来，而是让 AI 先把之前的语境信息压缩整理好，等下次启动时，只需要快速调取那些最必要的上下文就行。

几个关键指标相当能打：在 LongMemEval 基准测试里拿下了 89.8% 的分数，保存记忆时索引几乎是零延迟，而且全程不需要额外的 API Key 或向量数据库，无论本地部署还是云端使用都没问题。 ：https://github.com/moorcheh-ai/memanto…

moorcheh-ai/memanto

Source: https://github.com/moorcheh-ai/memanto

A companion memory agent that lets your agents focus and improve while you keep ownership of everything they learn.

Persistent memory for Claude Code, Cursor, Codex, and 14+ other agents, built on the world's first information-theoretic search engine. 100% free, open source, and runs entirely on your machine - no API keys, no vector database, no backend to babysit.

What you get

• No more re-explaining your codebase after every context reset. Memanto persists across sessions, your agent picks up where it left off.
• Fewer tokens burned on repeated context. Memories are retrieved only when relevant, so context windows go further.
• Memories searchable the instant they’re stored. Zero indexing wait, no LLM extraction tax at write time.
• One pip install. No vector DB to provision, no schema, no rerankers, no backend service to babysit.
• Flexible deployment. Choose between running the backend fully local, using it as a cloud SaaS, hosting in your own VPC, or switching between any of these options anytime you want.

Integrations

Works with Claude Code, Cursor, Codex, Windsurf, Cline, Continue, Goose, GitHub Copilot, and more. See the full list →

memanto connect <integration-tool-id> # integrates in one command
#eg: memanto connect claude-code    

The Six Gaps

Most memory tools are passive infrastructure — agents have to query them, parse the results, and figure out what to do. Memanto is an active memory agent built from the gaps models themselves named:

“My memory exists as a static snapshot injected into context — useful, but fundamentally passive.” — A model quote that became Memanto’s design brief.

Benchmarks

• 89.8% on LongMemEval and 87.1% on LoCoMo — outperforming Mem0, Zep, and Letta. Public datasets →
• Three primitives, not two: remember, recall, and answer LLM-grounded responses from memory, no extra API key.
• Single-query retrieval. No multi-stage pipelines, no graph schema, no rerankers.
• Typed semantic memory. 13 categories — instruction, fact, decision, goal, preference, relationship, and more.

Architecture

Memanto’s retrieval is powered by Moorcheh, an information-theoretic semantic engine. It runs as a local Docker container (free, no account) or as a free cloud service (100K free operations) the memanto CLI manages either for you.

On-Prem

Why Moorcheh?

Moorcheh is the semantic engine behind Memanto’s retrieval. Unlike vector databases that rely on approximate search and require indexing pipelines, Moorcheh uses an information-theoretic approach that returns exact results with zero indexing delay, write a memory and it’s searchable immediately.

This means Memanto doesn’t need a separate vector DB, embedding pipeline, or reranking stage. The Moorcheh engine runs as a local Docker container for on-prem users (no account needed) or as a managed cloud service with a free tier. Either way, it’s invisible - the memanto CLI handles it.

Setup & Demo

Local Dashboard For Best UX

CLI Reference

For a complete command reference, see the CLI User Guide.

Supported Memory Types

instruction, fact, decision, goal, commitment, preference, relationship, context, event, learning, observation, artifact, error

Use memory types to categorize what you store so retrieval is cleaner and more controllable:

• Save with a specific type: memanto remember "User prefers concise answers" --type preference
• Filter by type when searching: memanto recall "user communication style" --type preference

📦 SDKs

• TypeScript / Node.js — @moorcheh-ai/memanto — boots a local Memanto server via uvx and exposes an ergonomic Memanto client (remember / recall / answer).

REST API

Memanto exposes a session-based REST API for programmatic access. Start the server locally:

memanto serve

Full endpoint reference is available at docs.memanto.ai/api and at http://localhost:8000/docs when the server is running.

Research

Memanto: Typed Semantic Memory with Information-Theoretic Retrieval for Long-Horizon Agents

@misc{abtahi2026memantotypedsemanticmemory,
      title={Memanto: Typed Semantic Memory with Information-Theoretic Retrieval for Long-Horizon Agents}, 
      author={Seyed Moein Abtahi and Rasa Rahnema and Hetkumar Patel and Neel Patel and Majid Fekri and Tara Khani},
      year={2026},
      eprint={2604.22085},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2604.22085}, 
}

Support

Have questions or feedback? We’re here to help:

• Docs: https://docs.memanto.ai
• Discord: Join our Discord server
• Email: [email protected]
• X / Twitter: @moorcheh_ai

MIT License