Overview
Mem0 has 51,000 GitHub stars, $24M in funding, and is AWS's exclusive memory provider for Agent SDK. Zep is backed by a temporal knowledge graph on Neo4j, and Supermemory markets itself as a universal memory API. All share one belief: to remember, an agent must first extract key information from conversations, structure it, and store it in a vector or graph database.
Joshua Adler and Guy Zehavi (Sauron Labs) threw that belief away. Their system, True Memory, has no vector database, no graph store, no GPU, and no LLM extraction at ingestion. The entire system is a single SQLite file running on a plain CPU — and it beat them all.
Results: A 30-Point Slap
On LoCoMo (1,540 questions across 10 multi-session dialogues):
| System | Architecture | LoCoMo Accuracy | |--------|--------------|-----------------| | 🥇 EverMemOS | GPU embeddings + Neo4j graph | 94.5% | | 🥈 True Memory Pro | SQLite file, CPU | 93.0% | | 🥉 Zep | Temporal Knowledge Graph | ~71% | | 4 Supermemory | Universal memory API | 65.4% | | 5 Mem0 | Vector + graph + LLM extraction | 61.4% |
EverMemOS, an order of magnitude heavier, leads by only 1.5 points. Mem0 — the industry's brightest star — trails by 31.6 points.
The Wrong Foundation: Discarding Answers Before Querying
Extraction-based memory compresses raw conversations into structured "facts" at ingestion, discarding the original text. The paper's core insight:
> What is discarded before an unknown query can never be recovered at retrieval time.
It's like being asked to highlight key points of a document before knowing the exam questions — unmarked content is permanently lost. When a user later asks something not flagged as important, the answer is already gone.
Cognitive Science Backing
The design follows three scientific lineages:
- Bartlett (1932): memory is reconstructive, not a perfect copy — but in AI, schemas are engineer-hardcoded, not naturally emergent.
- Tulving (1972): episodic vs. semantic memory is decided at retrieval time, not storage. Encoding specificity: recall works best when retrieval cues match encoding conditions — extraction templates and future queries often speak different languages.
- Craik & Lockhart (1972): depth of processing couples encoding and retrieval across time.
The Killer Experiment: 92% of Errors Fixed
Researchers took 357 LoCoMo questions that an early True Memory version got wrong. They changed nothing about retrieval — just fed the full raw transcripts to the same answering model.
92% of the errors were fixed.
The "forgotten" information was always there. The agent didn't forget; the retrieval system simply failed to surface it. This refutes industry defenses: storage size, embedding quality, graph complexity, and extraction optimization are all beside the point.
Three Tiers of Architecture
| Tier | Systems | LoCoMo | Mechanism | |------|---------|--------|-----------| | 1: Extraction-based | Mem0, Supermemory | 61–65% | LLM extraction, raw text discarded | | 2: Retrieval baseline | BM25, Engram, RAG-ChromaDB | 80–86% | Raw text kept, similarity search only | | 3: Retrieval pipeline | True Memory | 89–93% | Raw text + multi-stage retrieval reasoning |
Tier gaps dwarf within-tier gaps: moving from discarding to keeping raw text gains ~20 points; adding retrieval reasoning gains another 7–12. Across 56 ablation configurations, accuracy spread within the top family was just 1.3% — architecture choice matters an order of magnitude more than component choice.
On BEAM-1M (1M-token conversations, 700 questions), True Memory Pro hit 76.6%, beating the previous best Hindsight (73.9%).
Uncomfortable Implications
1. The industry's "memory tax" may be unnecessary: Mem0 needs ≥2 LLM calls per write (~2s latency); True Memory ingests with zero LLM calls in milliseconds. 2. $24M may have backed the wrong assumption — that storage schema is the center of memory systems. 3. True Memory's encoding gate was disabled during benchmarks. It measures novelty via gzip compression cost (AUC 0.788 vs. 0.484 for cosine similarity baselines). Benchmarks reward "remember everything," so 93.0% was achieved handicapped — the gap could grow with better benchmarks.
The Bet
> Within two years, all mainstream agent memory frameworks will migrate toward retrieval-centered architectures: demoting storage to an event log and upgrading retrieval into the cognitive engine.
If wrong, it would mean cloud-native complexity carries irreducible engineering value. But a thirty-point gap doesn't look like a toy.
Paper Information
| Item | Detail | |------|--------| | Title | Storage Is Not Memory: A Retrieval-Centered Architecture for Agent Recall | | Authors | Joshua Adler, Guy Zehavi (Sauron Labs) | | arXiv | 2605.04897 | | Published | 2026-05-06, 17 pages | | Key benchmarks | LoCoMo 93.0% / LongMemEval 87.8% / BEAM-1M 76.6% |
*Based on arXiv:2605.04897. Rankings across systems are valid; absolute scores use a semantic-match judge and should not be directly compared to strict-match baselines.*