Agentic memory is evolving from ephemeral context-retrieval (RAG) into persistent, version-controlled, and bi-temporal ‘knowledge engines’ that treat an agent’s history as an auditable codebase.
Evidence
- Engram and G-Long introduce bi-temporal knowledge graph extraction and structured memory graphs to replace high-cost full-history context replay.
- GitOfThoughts and DeltaDB apply version-control primitives like diff, merge, and replay to agent reasoning and state management.
- Rosetta Memory and Infini Memory provide frameworks for decoupled, topic-structured memory that enables state portability across different model backbones.
Implications
- Builders must shift from simple vector search to managing stateful memory infrastructure that prioritizes provenance and long-term coherence.
- This transition enables more reliable long-horizon agents but introduces significant architectural complexity for state synchronization and auditability.
Concepts
Agent Memory Agents RAG Context Engineering Code Agents
Confidence
high