13 Tools Covered · Neutral · Updated April 2026

Memz

The AI Agent Memory Encyclopedia

Every memory tool. Compared honestly. Updated constantly. Memz is a neutral reference for developers building AI agents — not a product, not a pitch. Just the clearest map of the memory landscape available.

Explore the Landscape Full Comparison →

The Problem

What is AI agent memory?

By default, every AI agent conversation starts from zero. The agent has no recollection of past sessions, no knowledge of your preferences, and no ability to learn over time. That is fine for a single chat. It is a serious problem for agents running production workloads.

AI agent memory is the infrastructure that lets agents persist knowledge across sessions — who the user is, what has been tried before, what decisions were made, what the agent has learned. Without it, every session is a cold start. With it, agents compound.

The challenge: memory is not one problem. It includes storage, retrieval, deduplication, temporal reasoning, personalization, and multi-agent coordination. No single tool covers all of it. Memz maps every tool in the space so you can choose the right combination for your use case.

Agents forget

Every session starts blank. Decisions, preferences, and context evaporate between runs.

No personalization

Agents cannot learn who they are talking to. Every user interaction starts at zero.

No compounding

Knowledge built in one session is lost by the next. Agents never get smarter.

The Landscape

Every major player, in one place.

13 tools across 5 categories. Each has a distinct lane — understanding the category matters before comparing features.

Memory Server

Hindsight

Bank-scoped temporal memory server with the highest published LongMemEval score.

91.4% LongMemEval (highest published). BEAM temporal architecture. Bank-per-agent isolation. MIT license. Cloud or self-hosted.

Pluggable Memory Layer

Mem0

Pluggable memory layer with automatic fact extraction and deduplication.

49% LongMemEval. Auto-dedup of facts and preferences. Knowledge graph on Pro tier. $19–249/mo cloud. OSS core available. Known issue: AND queries broken across entity types.

Agent Runtime

Letta

Agent runtime with built-in 3-tier memory (formerly MemGPT).

Core / recall / archival memory tiers. Shared memory blocks for multi-agent. $20–200/mo. Pre-1.0. High lock-in — your agents live inside Letta's runtime.

Personalization Layer

Honcho

User personalization layer with dialectic reasoning. By Plastic Labs.

Cross-app user behavioral profiles. Dialectic reasoning over observed behavior. Open source. Not a memory store — a complementary layer that adds user modeling on top of any backend.

Memory Server

Zep

Long-term memory for AI assistants with knowledge graph extraction.

Knowledge graph extraction from conversations. Temporal awareness. Cloud and self-hosted options. Production-ready with paid tiers.

Vector Database

ChromaDB

Open-source vector database. Local-first. Zero latency.

Embedding storage and retrieval. Offline-capable. Simple Python API. Best choice for local dev or air-gapped setups. No agent memory management built in.

Vector Database

LanceDB

Serverless vector database in Rust. Hybrid search built in.

Lance columnar format. Vector + keyword hybrid search. Serverless — no infrastructure to manage. Production-ready. Self-hosted or cloud.

Vector Database

Supabase pgvector

PostgreSQL with vector extensions. SQL power meets semantic search.

1024-dim Jina v5 embeddings. Full SQL + vector similarity in one query. 100% data portability. Production-grade. You build all agent memory plumbing yourself.

Vector Database

Pinecone

Managed vector database for enterprise-scale retrieval.

Serverless architecture. Sub-millisecond queries at scale. Metadata filtering. No self-hosted option. Proprietary — cloud lock-in.

Vector Database

Weaviate

Vector and graph hybrid database. Multi-modal. Self-hosted or cloud.

Vector + graph hybrid. Multi-modal (text, image). Rich filtering. Self-hosted or Weaviate Cloud. Apache-2.0. Strong production track record.

Vector Database

Qdrant

High-performance vector search engine in Rust. Payload filtering.

Rust-powered. Dense + sparse vectors. Payload metadata filtering. Self-hosted or Qdrant Cloud. Apache-2.0. Strong performance benchmarks.

Orchestration Framework

LangGraph

Graph-based agent orchestration with checkpointing and state persistence.

State machine graph for agent flows. Checkpointing for resumable sessions. Built on LangChain. Cloud or self-hosted. Framework lock-in — your agent logic lives inside LangGraph.

Orchestration Framework

CrewAI

Role-based multi-agent framework with 4 built-in memory types.

Short-term, long-term, entity, and contextual memory. Role-based agent crews. Integrates with any LLM. MIT license. Opinionated framework — adapts poorly to custom memory backends.

Know what type of tool you need.

These are fundamentally different types of systems. Comparing a vector database to a memory server is like comparing a hard drive to a brain.

Memory Servers

Hindsight, Zep

Purpose-built systems for storing, retrieving, and reasoning over agent memories across sessions. They own the full lifecycle: write, deduplicate, recall.

Pluggable Memory Layers

Mem0, Honcho

Bolt-on layers that add memory capabilities on top of your existing stack. Mem0 handles fact extraction and dedup. Honcho handles user personalization.

Vector Databases

ChromaDB, LanceDB, Supabase pgvector, Pinecone, Weaviate, Qdrant

Store and retrieve embeddings by semantic similarity. Not memory systems by themselves — they are the storage layer. You provide the agent memory logic on top.

Agent Runtimes with Memory

Letta

Frameworks that embed memory directly into their agent execution model. High capability but high lock-in — your agents live inside the runtime.

Orchestration Frameworks

LangGraph, CrewAI

Multi-agent coordination systems with state persistence and checkpointing. Memory is a side effect of agent state, not a first-class feature.

Quick Compare

Six key dimensions at a glance.

Top 8 most-used tools. For the full 15-dimension breakdown, see the deep comparison.

Dimension	Hindsight	Mem0	Letta	Zep	ChromaDB	Supabase pgvector	Honcho	Qdrant
LongMemEval	91.4%	49%	—	—	—	—	—	—
Self-hosted	Yes	Partial	Yes	Yes	Yes	Yes	Partial	Yes
Auto-dedup	No	Yes	Partial	Partial	No	No	No	No
Temporal awareness	Yes	No	Partial	Yes	No	No	No	No
Semantic search	No	Yes	Yes	Yes	Yes	Yes	No	Yes
Open source	MIT	Partial	Apache-2.0	Apache-2.0	Apache-2.0	Apache-2.0	Yes	Apache-2.0

Based on published benchmarks and documented behavior — April 2026

Go deeper with the full comparison.

8 tools. 15 dimensions. Color-coded cells. Honest recommendations based on your actual use case — not vendor marketing.

Full Comparison →