Skip to content

Landscape

A factual survey of production agent memory systems, multi-agent frameworks, and protocols as of April 2026. This is reference material, not a competitive analysis.

Mem0 — Extraction-based, dual vector+graph, actor-aware memories. 66.9% LOCOMO benchmark.

Mastra — Observer+Reflector agents with text-only observations. 94.87% LongMemEval (highest recorded as of survey date).

Zep/Graphiti — Temporal knowledge graph with bi-temporal tracking and edge-based contradiction detection. 94.8% DMR.

Letta/MemGPT — LLM-as-OS model with 4-tier self-editing memory and sleep-time agents.

Cognee — KG+vector hybrid with 14 retrieval modes. Zero-infrastructure default (SQLite+LanceDB+Kuzu).

Hindsight — 4 biomimetic networks (world/bank/opinion/observation). 91.4% accuracy.

OMEGA — Local-first MCP server. 95.4% LongMemEval at 50ms latency.

Cog — Convention-based plain text memory. No code, no server, no runtime. Self-evolving rules via /evolve.

Memvid — Single-file portable memory with append-only smart frames.

LangGraph — Reducer-driven state with checkpointing and a Store interface for cross-thread memory.

CrewAI — Role-playing agents with unified Memory and configurable weighting.

AG2/AutoGen — Event-driven core with MemoryStream pub/sub and GroupChat coordination.

OpenAI Agents SDK — Lightweight handoffs with sessions for persistence. Memory is an acknowledged gap.

MCP — Universal tool transport (Linux Foundation). No memory semantics.

A2A — Agent discovery and task delegation. No shared memory semantics.

ESAA — Event sourcing for autonomous agents. Append-only activity log for task orchestration.

CoALA (2024) — Modular cognitive architecture framework.

A-MEM (NeurIPS 2025) — Zettelkasten-inspired bidirectional linking. 2x multi-hop reasoning improvement.

Multi-agent memory from computer architecture perspective (March 2026) — Identifies cache coherence for agents as a major open problem.

DLM (March 2026) — Three-system autonomous learning with learnable consolidation schedules.