zer0dex from Hermes Labs' LPCI research nails an insight that the rest of the agent memory space has overlooked: a structured index on top of vector search dramatically improves cross-domain retrieval.
Their dual-layer approach — a compressed markdown file (~782 tokens) acting as a semantic table of contents, plus a ChromaDB vector store — achieves 80% cross-reference accuracy where pure vector RAG scores 37.5%.
That's not a marginal improvement. That's a 2x gain on the hardest memory retrieval problem.
We respect the work. This article is an honest comparison.
The Two Systems
zer0dex: Elegant Minimalism
Two layers, one job:
- Compressed Index (MEMORY.md): Human-authored markdown (~3KB) with cross-domain pointers. Always loaded in context.
- Vector Store (mem0 + ChromaDB): Semantic fact database queried on every message (70ms). Top 5 results auto-injected.
AllSource Prime: Event-Sourced Memory Engine
One engine, three modalities:
- Events: Every memory mutation is an immutable, timestamped event. Full audit trail.
- Vectors: Semantic search via HNSW index, stored as projections over the event stream.
- Compressed Index: Auto-generated markdown organized by domain — not manually maintained.
Head-to-Head
Where zer0dex Wins Today
| Dimension | zer0dex | AllSource |
|---|---|---|
| Time to first memory | pip install + seed |
cargo install allsource-prime |
| Token overhead | 886 tokens/message | Variable |
| Python-native | Yes | SDKs available, core is Rust |
The compressed index is zer0dex's real contribution. Simple, cheap, and it works.
Where AllSource Wins
| Dimension | zer0dex | AllSource |
|---|---|---|
| Temporal reasoning | None | Full time-travel (as_of on every query) |
| Provenance | Mutable blobs | Immutable events with audit trail |
| Durability | ChromaDB (SQLite) | WAL (CRC32 checksums) + Parquet |
| Multi-agent | Single-machine | Multi-tenant, CRDT sync |
| Query performance | 70ms vector search | 12μs projection lookups |
| Index maintenance | Manual | Auto-generated projection |
| Scaling | Single Ollama instance | Leader-follower replication |
The Temporal Gap Is Critical
zer0dex has no concept of when a memory was true. If Alice was project lead until January and Bob took over in February, zer0dex returns both facts with no way to distinguish which is current.
AllSource stores every state change as an immutable event. Time-travel queries reconstruct the agent's knowledge at any point in history.
The Provenance Gap Matters Too
In AllSource, every memory mutation is an append-only event. You can trace any memory back to the event that created it, the session that triggered it, and the source material it was derived from.
Retrieval Quality
| System | Benchmark | Score |
|---|---|---|
| zer0dex | Custom (n=97) | 91.2% recall, 80% cross-ref |
| AllSource | Throughput-focused | 469K events/sec, 12μs query |
| Mem0 | LongMemEval | 49–66% |
| Letta | LoCoMo | ~83.2% |
| Zep | LoCoMo | ~85% |
The Compressed Index — Made Automatic
zer0dex's MEMORY.md is manually authored. As the knowledge base grows, keeping it accurate becomes a maintenance burden.
AllSource's CompressedIndexProjection auto-generates the equivalent markdown from graph events:
# Knowledge Index
_9 nodes, 3 domains, 2 cross-domain links_
## Domains
### revenue
- **Nodes:** 3
- **Types:** metric (2), decision (1)
- **Examples:** Q3 Revenue, Churn Rate
### engineering
- **Nodes:** 3
- **Types:** service (1), feature (1), metric (1)
## Cross-References
- **engineering ↔ revenue**: 2 edges (requires, improves)
- **product ↔ engineering**: 1 edge (depends_on)
No human maintenance. Updates incrementally as events flow in.
Getting Started
cargo install allsource-primeAdd to Claude Desktop config:
{
"mcpServers": {
"prime": {
"command": "allsource-prime",
"args": ["--data-dir", "~/.prime/memory"]
}
}
}13 MCP tools. Zero cloud dependencies. Full history.
The compressed index was the missing piece. Credit to Hermes Labs for proving it works.