A new paper from Sauron Labs — "Storage Is Not Memory: A Retrieval-Centered Architecture for Agent Recall" by Joshua Adler and Guy Zehavi — challenges the foundational assumption behind mainstream agent memory systems like Mem0, Zep, and Supermemory.
The Industry Consensus Being Challenged
Mem0 (51,000 GitHub stars, $24M funding, AWS's exclusive Agent SDK memory provider), Zep (temporal knowledge graph on Neo4j), and Supermemory (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, retrieving it when needed.
True Memory rejects this entirely: no vector database, no graph store, no GPU, no LLM extraction at ingestion. The whole system is a single SQLite file on a CPU — with a six-layer retrieval pipeline on top.
Benchmark Results
On LoCoMo (1,540 questions, 10 multi-session conversations):
| System | Architecture | LoCoMo Accuracy | |---|---|---| | EverMemOS | GPU embeddings + Neo4j graph | 94.5% | | True Memory Pro | SQLite file, CPU | 93.0% | | Zep | Temporal Knowledge Graph | ~71% | | Supermemory | Universal memory API | 65.4% | | Mem0 | Vector + graph + LLM extraction | 61.4% |
EverMemOS leads by only 1.5 points despite an order of magnitude more infrastructure. Mem0 trails by 31.6 points.
Other benchmarks: LongMemEval 87.8%, BEAM-1M 76.6% (1M-token conversations, 700 questions — above the previous best, Hindsight, at 73.9%).
The Core Argument
Extraction-based memory compresses conversations into "facts" at ingestion and discards the original text. The paper's central insight:
> Content discarded before a query is known can never be recovered at retrieval time.
If the LLM decides at ingestion what's "worth remembering," any future question outside that judgment is answered from permanently deleted context.
The Smoking Gun: 92% of Errors Fixed
The authors took 357 LoCoMo questions that an early version of True Memory answered incorrectly and fed the full raw conversation transcript directly to the same answer model — no retrieval changes, no tuning. 92% of errors disappeared.
The information was always there. Failures were retrieval failures, not storage failures — refuting the industry's standard defenses (bigger stores, better embedders, more LLM calls).
Three-Tier Structure
| Tier | Systems | LoCoMo | Mechanism | |---|---|---|---| | 1: Extraction-based | Mem0, Supermemory | 61–65% | Discard originals at ingestion | | 2: Retrieval baseline | BM25, Enragm/RAG-ChromaDB | 80–86% | Keep originals, simple similarity | | 3: Retrieval pipeline | True Memory | 89–93% | Keep originals + multi-stage retrieval reasoning |
The gap between tiers dwarfs the spread within tiers. A 56-configuration ablation found the top retrieval-centered family's accuracy spread was just 1.3% — component choices (embedder, reranker) matter far less than the architectural choice between extraction and retrieval, which swings results by 20+ points.
Cognitive Science Grounding
The design draws on:
- Bartlett (1932) — memory as reconstructive process; in AI, schemas are engineer-hardcoded, not emergent
- Tulving (1972) — encoding specificity: recall succeeds when retrieval cues match encoding conditions; extraction templates and future queries often speak different languages
- Craik & Lockhart (1972) — encoding depth determines retrievability; encoding and retrieval are coupled across time
- 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, 4 figures, 7 tables
True Memory stores only an event log (substrate); cognition happens at retrieval.
Uncomfortable Implications
1. The "memory tax" may be unnecessary. Mem0 requires ≥2 LLM calls per write (~2s latency); Zep needs dual Neo4j + vector infrastructure. True Memory ingests with zero LLM calls, millisecond latency. 2. The novelty gate was disabled during benchmarking. True Memory's encoding gate measures novelty via gzip compression cost (AUC 0.788 vs. 0.484 for cosine similarity) but was turned off because current benchmarks reward remembering everything. The 93.0% was achieved handicapped.
The Bet
The authors (and the reviewed post) predict that within two years, all mainstream agent memory frameworks will migrate toward retrieval-centered architectures: demoting storage to an event log and promoting retrieval to 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.