Built on Moorcheh · Open Source

MemantoMemory that AI agents love.

Open-source agentic memory built on Moorcheh.

13 typed memory categories · deploys with Moorcheh in your VPC

At a glance

89.8%
LongMemEval
<90ms
recall latency
10+
integrations
memanto cli · quick-demoStep 1 · Create agent
$
Create agent
Remember
Recall
Answer

Memanto agent quick-demo → Moorcheh namespace memanto_agent_quick-demo

The problem

LLMs forget between sessions

LLMs inherently lack persistent memory. Agents need a way to remember context, preferences, and decisions across conversations, sessions, and workflows.

  • Remember user preferences across sessions
  • Track decisions and commitments made during conversations
  • Store facts and context for faster, smarter responses
  • Manage long-running workflows with consistent state
  • Learn from interactions over time
session.state

Without Memanto

User: “Set theme to dark mode.”
— new session —
User: “Use my preferred theme.”
Agent: “Which theme do you prefer?”

With Memanto

recall → preference: dark mode
Agent: “Applying dark mode from your saved preference.”

Benchmark results

State-of-the-art memory accuracy

Evaluated on long-horizon agent benchmarks against leading memory systems.

LongMemEval

500 samples

Multi-session recall over long conversation horizons.

89.8%
+4.5% vs Mem0g
Memanto89.8%
Mem0g85.3%

LoCoMo

10 samples

Long-context conversational memory accuracy.

87.1%
+4.7% vs Zep
Memanto87.1%
Zep82.4%

Evaluation playground

Test the benchmarks yourself

Run the same LongMem and LoCoMo questions used in our published results. Pick a ground-truth prompt, send it through Memanto with your model, and score the answer with an LLM judge.

01

Pick a question

Choose from LongMem or LoCoMo ground-truth benchmarks.

02

Run inference

Query Memanto with your LLM and API key — nothing stored locally.

03

Judge accuracy

A secondary LLM scores the answer against ground truth.

Open playgroundRead research paper

API keys stay in your browser session only.

Why existing tools fail

Six gaps Memanto closes

Each gap maps to a production principle — built from real agent failure modes, not theoretical RAG limitations.

Gap 01

The problem

Irrelevant memory dumps

Memory arrives as one giant blob dumped into context. The agent cannot search, filter, or pull only what matters for the task.

Memanto closes it

Relevant results only

Instead of overloading your agent with data, Memanto finds only the exact information needed for the current task.

Gap 02

The problem

Outdated memories

A preference from six months ago carries the same weight as a deadline from yesterday. No timestamps, no recency signals.

Memanto closes it

Prioritizes new info

Memanto automatically keeps memories up to date by making sure new information always outranks older, stale notes.

Gap 03

The problem

Unknown memory sources

The agent cannot tell facts you stated from inferences it made, or data that has simply gone stale.

Memanto closes it

Verifiable sources

Memanto tracks where every memory came from, so your agent knows if a fact was told by you or just inferred from context.

Gap 04

The problem

All memories grouped together

Facts, habits, events, and instructions sit in one pile — no type labels, hierarchy, or filtering.

Memanto closes it

Smartly categorized

We organize memories into 13 clear types like facts and preferences, so agents only search for what is actually useful.

Gap 05

The problem

Memory contradiction

New information that conflicts with old memory is never resolved. Both versions live on side by side.

Memanto closes it

Resolves contradictions

Memanto catches and flags conflicting information immediately, preventing your agent from storing two different versions of the truth.

Gap 06

The problem

Long overhead ingestion

Traditional RAG needs indexing pipelines and latency before a memory is actually available for recall.

Memanto closes it

Instant memory ingestion

New memories are ready to use the second they are written. There is no waiting for indexing, and recall happens in less than 90ms.

Get started in 60 seconds
terminal
$ pip install memanto
$ memanto agent create quick-demo
[OK] Namespace created in Moorcheh: memanto_agent_quick-demo
Agent activated automatically.
$ memanto remember "I prefer responses in JSON format" --type preference
$ memanto recall "what response format, I prefer?"
$ memanto answer "what format should I use?"
13 Typed Memory Categories

Episodic, semantic, and procedural memory are no longer collapsed into one undifferentiated blob. Typed categories mean cleaner retrieval, better conflict detection, and controllable filtering.

instructionfactdecisiongoalcommitmentpreferencerelationshipcontexteventlearningobservationartifacterror
$memanto remember "I prefer responses in JSON format" --type preference
$memanto recall "what response format, I prefer?"

Results are backed by peer-reviewed research — Memanto outperforms Mem0, Zep, and Letta on long-horizon agent memory benchmarks.

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

Give your agents a brain that lastsOpen source on GitHub · pip install memanto