This post presents a technical analysis of the paper Storage Is Not Memory: A Retrieval-Centered Architecture for Agent Recall by Joshua Adler and Guy Zehavi (Sauron Labs), arXiv:2605.04897, published 2026-05-06.
Core Claim
The paper argues that an agent memory system's behavior is determined not by its storage schema but by the computation performed by the retrieval pipeline at query time — formally:
\(\text{Memory} \equiv \text{Retrieval}(\text{Query}, \text{Substrate})\), not \(\text{Memory} \equiv \text{Storage}(\text{Schema})\).
Key points
- Architecture: True Memory separates three phases — ingestion (raw events stored verbatim in a Messages table after an optional encoding gate), post-ingestion batch computation (embeddings, keywords, temporal markers, contradiction flags), and query-time retrieval across six layers: event log filter → messages table → similarity search → keyword augmentation → temporal resolution → ranking. All interpretive structure is computed at retrieval time; nothing is semantically discarded at ingestion.
- Three-tier evidence: Tier 1 systems (Mem0, Supermemory, LLM extraction at ingestion) score 61–65% on LoCoMo; Tier 2 (BM25, Enragm/RAG-ChromaDB, verbatim storage, pure similarity) score 80–86%; Tier 3 (True Memory Pro/Edge, verbatim storage plus multi-stage retrieval reasoning) score 89–93%. The gap between tiers (~20 pp and ~7–12 pp) far exceeds within-tier variation (1.3 pp).
- Bottleneck diagnosis: For 357 LoCoMo questions a earlier version answered incorrectly, replacing retrieval results with full conversation context — with the answering model fixed — repaired 92% of errors, showing retrieval quality, not storage capacity, is the dominant error source.
- Ablation: 56 configurations across 6 components showed a standard deviation of only 1.3 pp within the top-performing family, versus a 28–32 pp gap between Tier 1 and Tier 3 — an architecture-to-component effect ratio of roughly 20:1.
- Long-context validation: On BEAM-1M (~1M tokens), True Memory Pro scored 76.6% versus the previous best Hindsight at 73.9%, showing the retrieval-centered approach holds at long range.
- Cognitive science mapping: The design mirrors Bartlett's schema theory (memory is reconstructive), Tulving's episodic/semantic distinction, encoding specificity (extraction templates encode in a language mismatched to queries), and Craik & Lockhart's levels of processing.
Encoding gate details
The novelty gate uses compression cost rather than embeddings:
where gz is gzip level-6 and M is the concatenated nearest-neighbor memory text. It achieved AUC 0.788–0.816 in sweeps, versus 0.484 for cosine similarity, which counterintuitively flags noise as novel and important updates as redundant.
Infrastructure comparison
| Dimension | True Memory | Mem0 | Zep | |---|---|---|---| | Storage backend | Single SQLite file | Vector DB + Graph | Neo4j + Vector | | External dependencies | None | Qdrant/PostgreSQL/Neo4j | Neo4j/Chroma/Redis | | Ingestion LLM calls | 0 | ≥1 | ≥2 | | Ingestion latency | <10 ms | ~500 ms–2 s | ~3 s |
Limitations
1. The encoding gate (selective ingestion) was disabled in benchmarks; the 93.0% score reflects full-ingestion conditions, so its production impact is unmeasured. 2. BEAM-1M's 1M tokens fall short of the months-long conversation design goal; no validation at weeks-to-years scale. 3. All scores use a semantic-match judge; rankings across systems are valid, but absolute scores (e.g., 93.0%) are not directly comparable to strict-match baselines.
Conclusion
The post concludes that architecture choice (retrieval-centered vs. extraction-centered) affects accuracy an order of magnitude more than any component choice. Within an extraction-based framework, component tuning caps accuracy around 65%; within a retrieval-centered one, components are mere fine-tuning knobs.