Memanto is a memory agent for AI agents, and the README positions it as a "partner agent that manages the memory of other agents": deciding what should be kept, what conflicts, what expires, and who needs to know. It emphasizes that "persistence has been solved, but management has not"—various platforms and vector stores will save what agents write, but they will not proactively tell you that two agents now hold opposing views on the same matter, that an early preference is quietly overriding a recent decision, or that a newly launched agent is about to redo work that has already been completed and rolled back. Memanto's goal is to act as a "chief of staff," deciding what goes into the repository, monitoring access, resolving contradictions, discarding outdated content, and briefing each agent before it acts.
Purpose: To run as an independent agent alongside a cluster of agents, responsible for cross-platform memory governance. It is not a library to be called, but a second agent with autonomous judgment; each of its six behaviors has a corresponding command, and it is not a roadmap item. The documentation also states that a nightly daily cycle can be run via memanto schedule enable: organizing new memories, merging duplicates across agents, and flagging contradictions for human review.
Main capabilities:
- Semantic search and recall: Based on an information-theoretic semantic engine, provided with the package, and able to run as a local Docker container or as a free cloud service.
- Namespace management: Each agent has its own namespace by default, allowing precise configuration of what each agent should know, with production operations agents isolated from draft experiments.
- Text/vector operations and RAG: The retrieval engine is built in, with no need to separately prepare a vector store, embedding pipeline, reranker, schema migration, or backend maintenance.
- Typed memory: 13 memory categories—instruction, fact, decision, goal, commitment, preference, relationship, context, event, learning, observation, artifact, error—making recall filterable.
- Expiration policy: Memories are active by default until a policy retires them. Expiration is timestamped, auditable, and revocable; expired content still appears in recall marked with [EXPIRED] along with the reason. Policies are located at ~/.memanto/policies/<agent>.yaml, containing a retention table by type and naming rules, with the first matching rule taking precedence. Agents without a configured policy will not expire any memory.
- Traceability: Each belief carries confidence, source, provenance, and timePoke to trace back when this belief entered the cluster.
- Scheduled tasks: memanto schedule enable runs organizing and cleanup daily.
- Export and migration: memanto memory export --okf exports to Open Knowledge Format (plain Markdown, readable, diffable, committable, greppable); memanto migrate can import from Mem0, Letta, Supermemory, Zep, Hindsight, or any OKF package, and the same command can be used in reverse.
- Backend switching: memanto config backend switches between local, free cloud, and self-hosted, and data migrates accordingly.
- Local dashboard: memanto ui, which supports browsing, searching, and auditing the entire memory asset.
Usage:
- Fully local: no account or API key required, data never leaves the machine, requires Docker, works with Ollama.
- Free cloud: no credit card required, about 60 seconds, get a free key at console.moorcheh.ai/api-keys, includes 100,000 free operations.
- Installation and integration: pip install; memanto connect claude-code connects to Claude Code, and also supports Cursor, Codex, Windsurf, Cline, Continue, Goose, Copilot, etc., one command each.
- Runtimes: supports an in-process Python client; TypeScript/Node.js's @moorcheh-ai/memanto starts a local Memanto server via uvx and exposes the remember/recall/answer client; Go connects via github.com/moorcheh-ai/memanto/sdks/go or starts the server via uvx; the REST API is started via memanto serve, with runtime endpoints referenced at docs.memanto.ai/api and http://localhost:8000/docs。
- Other commands: memanto daily-summary provides a readable summary; memanto status shows registered agents, sessions, and health status; memanto memory restore can restore memory.
Applicable scenarios: teams and organizations that need to manage agent memory across platforms, handle multi-agent contradictions, control memory expiration, and require data sovereignty and auditability; suitable for users who want to avoid lock-in by platforms' native memory features and need a portable OKF format.
Limitations and notes: the documentation states that on public recall benchmarks it reports LongMemEval 89.8% and LoCoMo 87.1%, with datasets and testing tools open-sourced at huggingface.co/moorcheh. It also explicitly reminds readers: benchmark scores across projects are not directly comparable, and the reader model, judge model, judge prompt, and retrieval budget can each cause fluctuations of several points; different public runs use different configurations, and all numbers (including its own) should be regarded as directional references; what a memory agent should ultimately measure is not recall rate, but long-term asset quality (contradiction rate, staleness, accuracy in the sixth month). The project uses the MIT license, with no open core tier, feature flags, or seat limits. The platform supports macOS, Linux, and Windows. Contact support@moorcheh.ai, @moorcheh_ai. Other details such as specific pricing and more integration details are not yet provided.
