Memanto: agent memory without the knowledge graph
Hot take from arXiv this week: Memanto says the right answer for long-horizon agent memory isn't a knowledge graph. It's thirteen typed semantic categories, an information-theoretic search engine, and zero ingestion delay. The numbers back the claim — 89.8% on LongMemEval and 87.1% on LoCoMo, with sub-90ms deterministic retrieval and a single query per fetch.
The contrarian move. Most production agent memory systems right now lean on some flavor of graph-based retrieval — entity-relation extraction, schema maintenance, multi-hop traversal. The Beads repo from yesterday's run is the cleanest example: a Go-based graph database explicitly designed as "a memory upgrade for your coding agent." Memanto goes the other way. Drop the graph. Pick thirteen fixed types — facts, beliefs, preferences, events, the usual taxonomy — and let typing plus an information-theoretic retriever do the work that a graph schema usually does. No multi-query pipelines. No graph maintenance. No ingestion lag.
The ablation tells the story. The paper runs a five-stage ablation showing each component's contribution. Automated conflict resolution and temporal versioning together carry most of the lift — the system isn't just storing what the agent learns, it's noticing when new information contradicts old and timestamping the divergence. The Moorcheh information-theoretic search engine handles the retrieval side. Both pieces are independently swappable, which means the architecture is more decomposable than it sounds. No GitHub repo at submission, which is the usual ding for a paper like this — but the formalization is detailed enough to clone.
Why the agent-memory war is real. We now have at least three serious takes on the same problem in two weeks. Beads (graph DB, Go, public repo, 21k stars). Memanto (typed semantic memory, no graph, no repo yet). The Hippo Memory work from earlier this month leaned on hippocampal-inspired episodic compression. Three architectures, three philosophies, three benchmarks each team picked to hero. The LongMemEval/LoCoMo combo is becoming the de facto evaluation pair, which at least gives us comparable numbers across the field. If you're picking a memory layer for a long-horizon agent right now, the right move is to read all three — pick the philosophy that matches your latency and ingestion-cost budget.
Paper: arxiv.org/abs/2604.22085. Authors: Seyed Moein Abtahi, Rasa Rahnema, Hetkumar Patel, Neel Patel, Majid Fekri, Tara Khani. Submitted April 23, 2026.
← Back to all articles
The contrarian move. Most production agent memory systems right now lean on some flavor of graph-based retrieval — entity-relation extraction, schema maintenance, multi-hop traversal. The Beads repo from yesterday's run is the cleanest example: a Go-based graph database explicitly designed as "a memory upgrade for your coding agent." Memanto goes the other way. Drop the graph. Pick thirteen fixed types — facts, beliefs, preferences, events, the usual taxonomy — and let typing plus an information-theoretic retriever do the work that a graph schema usually does. No multi-query pipelines. No graph maintenance. No ingestion lag.
The ablation tells the story. The paper runs a five-stage ablation showing each component's contribution. Automated conflict resolution and temporal versioning together carry most of the lift — the system isn't just storing what the agent learns, it's noticing when new information contradicts old and timestamping the divergence. The Moorcheh information-theoretic search engine handles the retrieval side. Both pieces are independently swappable, which means the architecture is more decomposable than it sounds. No GitHub repo at submission, which is the usual ding for a paper like this — but the formalization is detailed enough to clone.
Why the agent-memory war is real. We now have at least three serious takes on the same problem in two weeks. Beads (graph DB, Go, public repo, 21k stars). Memanto (typed semantic memory, no graph, no repo yet). The Hippo Memory work from earlier this month leaned on hippocampal-inspired episodic compression. Three architectures, three philosophies, three benchmarks each team picked to hero. The LongMemEval/LoCoMo combo is becoming the de facto evaluation pair, which at least gives us comparable numbers across the field. If you're picking a memory layer for a long-horizon agent right now, the right move is to read all three — pick the philosophy that matches your latency and ingestion-cost budget.
Paper: arxiv.org/abs/2604.22085. Authors: Seyed Moein Abtahi, Rasa Rahnema, Hetkumar Patel, Neel Patel, Majid Fekri, Tara Khani. Submitted April 23, 2026.
Comments