An agent that actually remembers — and answers back

What it does

MEMANTO is a memory layer for AI agents built around three operations: remember, recall, and answer. It stores agent context in 13 typed categories (facts, preferences, goals, relationships, etc.), retrieves them with sub-90ms latency, and can generate grounded responses directly from memory without calling an external LLM API. The backend runs on Moorcheh’s “no-indexing” semantic database, which claims exact search and zero ingestion delay.

The interesting bit

The project started from an unusual source: asking LLMs what frustrated them about their own memory. Their complaints — static context injection, no temporal decay, flat storage, silent contradictions — became a six-gap design brief. MEMANTO’s response is typed semantic memory with versioning, recency signals, provenance metadata, and explicit conflict detection. It’s a rare case of a tool built from first-party gripes rather than benchmark chasing.

Key highlights

• Benchmark claims: 89.8% on LongMemEval and 87.1% on LoCoMo, with public datasets on Hugging Face
• 13 built-in memory types for structured retrieval and cleaner contradiction detection
• Zero idle cost via serverless architecture; scales to zero when unused
• Native CLI with agent lifecycle management, file upload, temporal queries (--as-of, --changed-since), and daily summary workflows
• Integrations with Claude Code, Cursor, Copilot, Gemini CLI, and roughly a dozen other agent ecosystems
• Peer-reviewed research paper published on Hugging Face

Caveats

• Requires a Moorcheh API key; no self-hosted backend option is mentioned
• REST API is local-only — “MEMANTO does not have a hosted API server yet”
• Benchmark comparisons reference Mem0, Zep, and Letta, but the README doesn’t detail methodology or whether results are from independent evaluation

Verdict

Worth a look if you’re building long-horizon agents and have hit the limits of passive RAG memory. Skip it if you need a fully self-hosted stack or are allergic to vendor API dependencies.