memanto 0.2.15

Memory that AI Agents Love!

Memanto is a Memory Agent; a companion agent that manages the memories of your other agents:
what to keep, what conflicts, what expires, and who needs to know.

pip install memanto

Every platform will store your agents' memory. None of them will manage it. Managing it across platforms is against their interest. That is the job of a Memory Agent.

Persistence is solved. Claude, Bedrock, Cursor, and every vector store will happily keep what your agents write. None of them will tell you that two of your agents now believe opposite things about your auth service, that a preference from March has quietly outranked a decision from last week, or that the agent you spun up this morning is about to redo work another one finished and reverted.

Storage is a filing cabinet. Memanto is the chief of staff; it decides what goes in, watches the access, resolves what contradicts, discards what's stale, and briefs each agent before it acts.

It's an agent, not an API

Memanto isn't a library you call. It's a second agent that runs beside your fleet and does six things on its own judgment. Each one is a real behavior with a command behind it, nothing here is a roadmap item.

And it does this while you're asleep. memanto schedule enable and the loop runs daily: new memories curated, duplicates merged across agents, contradictions flagged for your review. You come back to a fleet that knows more than it did yesterday, without having sorted anything yourself.

60 seconds to a managed fleet

pip install memanto
memanto                            # "On-Prem" (Docker, no account) or "Cloud" (free key)
memanto connect claude-code        # also: cursor, codex, windsurf, cline, goose, copilot…

Your agents now share one managed estate. No code changes, no wrapper, no rewrite of your agent loop.

# backend-agent learns something on Monday
memanto remember "Auth migrated to JWT — session cookies deprecated" --type decision

# review-agent, which never saw that session, knows it on Friday
memanto recall "how does auth work"
memanto answer  "why did we drop session cookies?"     # grounded, no extra API key

# what did the fleet believe last Tuesday? what changed since the release?
memanto recall "deployment policy" --as-of 2026-08-05
memanto recall "deployment policy" --changed-since v2.1

macOS, Linux, Windows. memanto ui opens a local dashboard over the whole estate — browse it, search it, audit it.

Own your agentic memory

This is the part that matters in two years, and it's the part every platform-native memory feature is designed to prevent.

Your estate is a file. memanto memory export --okf gives you the Open Knowledge Format — plain Markdown, readable, diffable, committable, greppable. Not a proprietary dump you can technically request. The actual working format.

It moves. memanto migrate imports from Mem0, Letta, Supermemory, or any OKF bundle. The same command works in reverse. OKF is an open interchange format any framework or vendor can implement — including ours' competitors, deliberately.

It runs on your machine. Local Docker + Ollama, no account, no API key, nothing leaves your infrastructure. Or free cloud, or your own hosting. memanto config backend switches between them in one command, and the estate comes with you.

MIT. No open-core tier waiting to gate the useful half. No feature flags, no seat limits, no rug pull.

There is no lock-in because there is nothing to lock.

Security & sovereignty

Nothing leaves your machine in on-prem mode. Docker + Ollama, no account, no outbound calls. The full loop — extraction, consolidation, reconciliation, briefing — runs locally.

Scoped by default. Each agent gets its own namespace. Your production-ops agent doesn't read your scratch experiments; you provision exactly what each one should know and nothing more.

Every belief is traceable. Confidence score, source, timestamp, and what it superseded. When an agent acts on something, you can walk back to where that belief entered the fleet and when — which is the difference between an auditable estate and a black box.

Unlike memory storage

Storage substrates sit beneath Memanto — vector stores, filesystems, and platform-native memory features are all backends it manages. Their commoditization is good for you: it makes the substrate free and leaves the management to something that's actually good at it.

⭐ Star the repo if Memanto is managing your fleet's memory
_{It's the signal that tells us to keep building this in the open, under MIT, with nothing held back.}

Developer experience

One pip install. No vector store to provision, no embedding pipeline, no reranker, no schema migration, no backend to babysit. The retrieval engine ships in the box.

Works with what you already run. memanto connect claude-code — same for Cursor, Codex, Windsurf, Cline, Continue, Goose, Copilot, and more. One command each.

Searchable the moment it's written. No extraction pass at write time, no graph to rebuild, no indexing queue. remember returns and every agent in the fleet can already recall it.

Typed, not soup. 13 memory categories — instruction, fact, decision, goal, preference, relationship, and more — so recall is filterable instead of one undifferentiated blob.

A dashboard, not a log file. memanto ui for the whole estate. memanto daily-summary for a readable digest of what changed across your agents. memanto status for registered agents, sessions, and health.

memanto remember "User prefers concise answers" --type preference
memanto recall "user communication style" --type preference

pip install memanto
memanto           # choose "On-Prem" — guides through Docker + Ollama setup

pip install memanto
memanto           # choose "Cloud" — paste your free API key

Watch it work

Research

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

On public recall benchmarks we report 89.8% on LongMemEval and 87.1% on LoCoMo. Datasets and harness are open at huggingface.co/moorcheh — run them yourself.

A caveat we'd rather say out loud: cross-project scores on these benchmarks are not comparable. Reader model, judge model, judge prompt, and retrieval budget each move results by several points, and no two published runs share a configuration. Treat every number in this category — including ours — as directional. What a Memory Agent should eventually be measured on isn't recall at all, but estate quality over time: contradiction rate, staleness, and precision at month six.

@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},
}

Community

Questions: support@moorcheh.ai · @moorcheh_ai

MIT License
_{English · Español · 简体中文 · 日本語}