memory-systems

Start with file-system memory storing facts as structured JSON with timestamps, then escalate to a vector store like Mem0 when semantic search is needed. Add a temporal knowledge graph such as Zep/Graphiti only when the agent requires relationship traversal or time-travel queries.

Choose Mem0 for fast time-to-production with managed infrastructure, Zep/Graphiti for bi-temporal modeling and relationship reasoning, and Cognee for dense multi-layer semantic graphs with customizable ingestion pipelines. Letta fits agents needing deep self-introspection over their own memory.

A temporal knowledge graph stores facts with valid_from and valid_until timestamps so agents can answer point-in-time queries like where a user lived on a specific date. It prevents stale facts from poisoning context by invalidating outdated relationships without discarding history.

Yes, in some benchmarks. Letta's filesystem-based agents scored 74% on LoCoMo using basic file operations, beating Mem0's specialized tools at 68.5%, which shows reliable retrieval matters more than tool sophistication for many workloads.

Unbounded memory growth, stale facts without validity tracking, and embedding model mismatches all degrade retrieval quality. Run periodic consolidation triggered by memory count thresholds, pin one embedding model per store, and apply expiry policies to aged facts.

Avoid graph structures during prototyping or when queries are simple factual lookups, since graphs add infrastructure cost and schema maintenance. Start with file-system or vector storage and add graph layers only when multi-hop reasoning or relationship traversal is demonstrably required.