archon-memory-core vs Zep
Zep is a temporal knowledge graph for LLM agents — backed by Graphiti, it models facts as nodes and edges with valid-time and transaction-time. archon-memory-core is a lighter retrieval library focused on contradiction resolution. Different tools for different shapes of memory.
Architectural difference
Zep runs as a Postgres-backed service. Ingested text is chunked, entities and relationships are extracted, and the result lands in a temporal knowledge graph. You query by natural language or by graph traversal, and Zep reasons about "as of time T." This is powerful — and ops-heavy. Community edition is self-hostable; Zep Cloud is managed.
archon-memory-core is an in-process Python library. Memory is flat rows with persistence class, priority, and embedding. There is no entity extraction, no graph, no server. Contradictions aren't modeled as temporal edges — they're allowed to coexist and the retriever picks the winner based on persistence and priority.
Feature comparison
| Capability | Zep | archon-memory-core |
|---|---|---|
| License | Apache 2.0 (community) | Apache 2.0 |
| Deployment | Postgres-backed server | In-process library |
| Data model | Temporal knowledge graph | Flat rows + embeddings |
| Entity extraction | Built-in (LLM-backed) | Out of scope |
| Temporal queries | First-class (valid-time, tx-time) | Timestamps only |
| Contradiction handling | New edges supersede old in graph | Retriever scores persistence + priority |
| Benchmark | DMR, LongMemEval | AMB v2.3 (99.2% top-1) |
| Install footprint | Docker + Postgres | pip install |
When to pick Zep
- You need time-travel queries: "what did the agent know about this user on March 1?"
- Your memory is fundamentally relational — people, companies, events, roles.
- You want entity extraction and graph construction handled for you.
- You're okay running another Postgres-backed service.
When to pick archon-memory-core
- Your memory is closer to "facts about the user" than "a knowledge graph of the world."
- You want contradictions surfaced by ranking, not by graph supersession edges.
- You don't want another service to deploy, monitor, and upgrade.
- Retrieval latency matters and you don't want graph traversal on the hot path.
When to use both
They aren't mutually exclusive. A real assistant might use Zep for the temporal world-model (organization charts, event timelines, longitudinal facts) and archon-memory-core for ephemeral and session-scoped facts that need rapid contradiction resolution without graph maintenance overhead.