Zero Memory: Provenance-Aware Long-Term Memory
Abstract
Large language model (LLM) assistants need durable, faithful memory of everything a user or organization has said and stored, yet most memory systems commit to a single organizing structure, a fact store, a vector index, or a knowledge graph, and inherit its blind spots. We present Zero Memory, a provenance-aware long-term memory system that distills a user's conversations, files, and connected sources into three parallel memory systems, each capturing a different facet of the same history: an episodic Memory Events timeline that makes when and what changed first-class, an associative entity-event knowledge graph that links people and projects across sessions, and a semantic, curated, citation-locked Hierarchical Documentary Memory of durable facts. A retrieval turn runs an intent gate, a source router, and then three concurrent searches, one per system, each a tool-using loop over hybrid embedding and lexical search under controlled filters; their grounded, cited answers are integrated into one answer with a single confidence. Every learned item carries its origin, timestamp, and evidence pointer, and every answer is read under a citation lock: it may cite only evidence its reader actually opened, so fabrication is structurally excluded and the system abstains rather than guesses. On two public benchmarks the system sets a new state of the art: 95.60% on LongMemEval and 93.60% on LoCoMo, improving over the strongest prior systems by 0.73 and 1.10 points. A controlled study across eight backbone LLMs shows accuracy varies by only 3.4 points while per-query cost varies by around 30x, with near-state-of-the-art quality at up to 20x lower cost per query, the signature of memory-driven, rather than model-driven, quality.
Results
On the two most widely used public benchmarks for long-term conversational memory, Zero Memory sets a new state of the art, and does so robustly across eight different backbone models, evidence that the gains come from the memory substrate rather than any single model.
LongMemEval (500 questions)
| # | System | Accuracy % |
|---|---|---|
| 1 | Zero (ours) | 95.60 |
| 2 | Mastra | 94.87 |
| 3 | Hindsight | 91.40 |
| 4 | EmergenceMem | 86.00 |
| 5 | Supermemory | 85.20 |
| 6 | Zep | 71.20 |
LoCoMo (1,540 questions)
| # | System | Accuracy % |
|---|---|---|
| 1 | Zero (ours) | 93.60 |
| 2 | Mem0 | 92.50 |
| 3 | ByteRover 2.0 | 92.20 |
| 4 | Hindsight | 89.60 |
| 5 | Memobase | 75.80 |
| 6 | Zep | 75.10 |
How it works
Most memory systems pick one shape, a fact store, a vector index, or a knowledge graph, and inherit its blind spots. Zero Memory keeps three memories at once, each tuned to a different kind of question.
- An events timeline. A merged, cross-source record of what happened and when, so the system can tell a fact from its later correction and answer "what did we decide, and when did that change?"
- An entity graph. People, projects, and events linked together, so questions that span months and conversations are answered by walking the connections rather than hoping for a keyword match.
- A curated documentary memory. A citation-locked core of durable facts, your profile, preferences, and the people you work with, written from your own material rather than guessed.
A retrieval turn runs a fast intent gate (self-contained turns skip memory entirely, adding no latency), routes to the right stores, and runs three searches at once, each a tool-using loop over hybrid meaning-and-keyword search. Their answers are merged into one, with a single confidence.
The reading discipline is what makes it trustworthy. Every item carries its origin, timestamp, and a pointer to the evidence. Every answer is read under a citation lock: it can cite only evidence it actually opened, so the system abstains instead of fabricating when it cannot support a claim. That is why an operator can inspect why the memory believes what it reports, and why quality holds across models rather than depending on one.