memory-systems

Start with file-system memory storing facts as structured JSON with timestamps, then move to a vector store like Mem0 when you need semantic search. Add a temporal knowledge graph such as Zep/Graphiti only when you need relationship traversal or time-based queries.

Mem0 offers the fastest path to production with managed infrastructure and broad integrations. Zep/Graphiti fits enterprise use cases needing temporal reasoning over relationships, while Letta provides the deepest agent self-editing and introspection through tiered memory storage.

A temporal knowledge graph stores facts as relationships with valid_from and valid_until timestamps, letting agents query what was true at a specific point in time. This prevents stale facts from poisoning context when information changes over time.

A vector store is enough for direct factual queries, but retrieval degrades on multi-hop reasoning. Benchmarks show Letta's simple filesystem agents scored 74% on LoCoMo, so add graph structure only when retrieval quality demands it.

Prefer the fact with the most recent valid_from timestamp and surface conflicts to the user when confidence is low. Consolidate memories periodically by invalidating outdated facts rather than deleting them, preserving history for temporal queries.

Empty results usually mean overly narrow filters, so fall back to broader searches by removing entity filters or widening time ranges. Stale results indicate expired valid_until timestamps, which should trigger memory consolidation before retrying.