Memanto offers open-source, local long-term memory for AI agents

Summary

Memanto is an open-source companion agent that manages long-term memory for AI agents across sessions. It is designed to collect important insights, consolidate context, and provide a briefing when an agent starts, with support advertised for Claude Code, Cursor, Codex, and more than 20 other agents. The project says memories can be converted between a semantic backend and human-readable Markdown files in an LLM Wiki format, allowing users to inspect, export, or migrate their data. Its command-line interface supports agent creation, remembering, recalling, answering questions from stored context, connecting integrations, and starting a local dashboard and REST API. The site also lists semantic search, built-in retrieval-augmented generation, multi-agent namespaces, semantic memory categories, conflict resolution, freshness handling, source tracking, temporal queries, confidence scoring, and daily summaries. Memanto can run on a user’s machine with Docker; its on-prem setup uses local Ollama models for embeddings and answers, although the site also says users can bring OpenAI or Cohere. The product page claims memories become searchable in under 90 milliseconds and highlights 30 integrations, 32x compression, and a write path with no LLM token cost, but these are presented as product claims rather than independently verified results. It says the project is MIT-licensed, free for on-prem use, and has no API key or usage cap. A managed Moorcheh Cloud option is described as free to start with about 100,000 operations, while hosted Memanto itself is presented as not yet built. The page also claims leading results on the LoCoMo and LongMemEval agent-memory benchmarks, without providing detailed methodology in the extracted material.