Give your AI agent
perfect memory
Knowledge graph + vector search + compressed index in one binary. No cloud dependency. No monthly fees. 12μs queries.
30 seconds to persistent memory
Add to ~/.claude/claude_desktop_config.json:
{
"mcpServers": {
"prime": {
"command": "allsource-prime",
"args": ["--data-dir", "~/.prime/memory"]
}
}
}13 MCP tools: graph CRUD, vector search, hybrid recall, compressed index, temporal queries.
How it works
How does an agent remember?
Every fact is an immutable event stored in a durable WAL + Parquet engine. Knowledge graph nodes, vector embeddings, and relationships are all events — giving you full history and time-travel for free.
- Graph nodes: entities, concepts, decisions
- Vector embeddings: semantic similarity search
- Compressed index: navigational scaffolding
How fast is recall?
12μs projection lookups via DashMap. Hybrid recall combines vector similarity, graph traversal, and temporal recency in a single query.
- Vector search: HNSW index over embeddings
- Graph expansion: 1-hop BFS from matches
- Compressed index: cross-domain reasoning
What about cross-domain questions?
The compressed index — an auto-generated markdown summary organized by domain — bridges the gap that pure vector search misses. When you ask 'how does X relate to Y?', the index provides cross-domain pointers that double retrieval accuracy.
- Auto-generated from graph events
- Organized by domain with cross-references
- 80%+ cross-domain recall accuracy
Can I time-travel?
Every mutation is an append-only event. Query any entity's state at any past timestamp. See who added what, when, and why.
- as_of queries: reconstruct past state
- Full audit trail: every create, update, delete
- Graph diff: what changed between two timestamps
How we compare
vs. zer0dex, Mem0, Letta, and Zep
| Feature | zer0dex | Mem0 | Letta | Zep | Prime |
|---|---|---|---|---|---|
| Compressed Index | Manual | No | No | No | Auto-generated |
| Temporal Queries | No | No | No | Yes | Yes |
| Provenance | No | No | Partial | Partial | Full event audit |
| Cross-Domain Recall | 80% | ~50% | ~37% | ~85% | 80%+ |
| Offline/Embedded | Yes | No | No | Optional | Yes + sync |
| Latency | 70ms | Variable | Variable | Variable | 12μs |
| Cost | $0 | $0-249/mo | Cloud | Cloud | $0 |
Sources: zer0dex · Vectorize 2026 Comparison
Use cases
Personal AI Assistant
Claude remembers your project context across sessions. Yesterday's decisions inform today's answers.
Multi-Agent Knowledge Sharing
Three agents work on different parts of a codebase. Findings flow through the shared graph.
Incident Response Memory
Your oncall agent remembers every past incident. 'What happened last time this alert fired?'
Research Assistant
Read 50 papers, build a knowledge graph. The compressed index surfaces unexpected cross-domain connections.
Code Review Context
Agent remembers past review feedback. 'Last time you said X about error handling in this module.'
Audit & Compliance
Full provenance on every memory. Who added what, when, from what source. Time-travel to any past state.
One engine, not three databases
┌─────────────────────────────────────────────┐ │ AllSource Prime │ │ │ │ Graph Vectors Temporal Compressed │ │ Nodes HNSW History Index │ │ Edges Embed Time-travel Cross-refs │ │ Similar Diff │ │ │ │ ┌──────────────────────────────────────┐ │ │ │ AllSource Core Engine │ │ │ │ WAL + Parquet + DashMap + HLC + CRDT │ │ │ │ 469K events/sec │ 12μs queries │ │ │ └──────────────────────────────────────┘ │ └─────────────────────────────────────────────┘
Other agent memory frameworks glue together a vector DB + graph DB + event store. Prime is one engine where vectors, graph nodes, and edges are all events in the same durable WAL.
Start remembering
One command. No cloud account. No API key.