8 Tools · 15 Dimensions · No Spin

AI Memory Tools Compared

An honest, technical head-to-head across the tools that matter most.

No tool is highlighted. No tool is the winner. Every tool has a lane where it excels and limits where it stops. Read this before you decide.

Based on published benchmarks and documented behavior — April 2026

Tool Profiles

Know what you are comparing.

Each tool occupies a different part of the memory stack. Understanding the lane matters before looking at the matrix.

Memory Server

Hindsight

Best for

Agents that need the highest published memory accuracy and temporal reasoning without building a full multi-backend stack.

Watch out for

No embedding search, no deduplication, no user modeling built in. You add those from other tools.

Pluggable Memory Layer

Mem0

Best for

Quick bolt-on memory with automatic fact extraction and deduplication. Good for product teams who want memory without building from scratch.

Watch out for

AND queries broken across entity types (issue #3218). Limited true multi-agent coordination. Only 49% LongMemEval.

Agent Runtime

Letta

Best for

Building new agents from scratch inside a managed runtime with built-in 3-tier memory (core / recall / archival).

Watch out for

Pre-1.0. High lock-in — your agents live inside Letta's runtime. No published benchmarks. Hard to migrate out.

Personalization Layer

Honcho

Best for

Consumer-facing agents that need persistent user behavioral profiles and cross-app identity. By Plastic Labs.

Watch out for

Not a memory backend — a complementary personalization layer. Cannot replace storage. Must be combined with a vector DB.

Memory Server

Zep

Best for

AI assistants that need long-term memory with knowledge graph extraction from conversations. Production-ready.

Watch out for

Cloud-first pricing can escalate. Knowledge graph quality depends on conversation structure.

Vector Database

ChromaDB

Best for

Local development, prototyping, and air-gapped setups where you need semantic search with zero infrastructure overhead.

Watch out for

Not a memory system. You build all the agent memory logic yourself. Not designed for production scale.

Vector Database

Supabase pgvector

Best for

RAG workloads where you need semantic search combined with full SQL query power and 100% data portability.

Watch out for

No agent memory management built in. No temporal awareness. You provide all the plumbing.

Vector Database

LanceDB

Best for

Production semantic search with hybrid (vector + keyword) retrieval. Serverless — no infrastructure to manage.

Watch out for

No agent memory management. Community smaller than Chroma or Supabase. Still maturing.

Feature Matrix

15 dimensions. 8 tools. No spin.

Green = strong. Amber = partial or conditions apply. Red = not available. No tool is highlighted — this is a neutral reference.

Dimension	Hindsight	Mem0	Letta	Honcho	Zep	ChromaDB	Supabase pgvector	LanceDB
Category	Memory Server	Pluggable Layer	Agent Runtime	Personalization	Memory Server	Vector DB	Vector DB	Vector DB
Open source license	MIT	Partial OSS core, cloud proprietary	Apache-2.0	Yes Open source	Apache-2.0	Apache-2.0	Apache-2.0	Apache-2.0
Self-hosted	Yes	Partial OSS limited vs cloud	Yes Apache-2.0	Partial Hosted by Plastic Labs	Yes	Yes Local-first	Yes Docker	Yes Serverless local
Cloud option + pricing	Yes Free tier + paid	Yes $19–249/mo	Yes $20–200/mo	Yes Cloud only, contact	Yes Paid tiers	Yes Chroma Cloud	Yes Free + paid tiers	Yes LanceDB Cloud
Memory persistence (cross-session)	Yes Bank-scoped, temporal	Yes Auto-extracted facts	Yes 3-tier model	Partial User profiles only	Yes Conversation + graph	Partial Embeddings only, no agent logic	Partial SQL rows persist, no memory mgmt	Partial Vectors persist, no memory mgmt
Multi-agent support	Partial Separate banks, no shared layer	Partial Broken AND queries #3218	Yes Shared memory blocks	Partial Cross-app user identity only	Partial User/session scoped	No No concept of agents	Partial RLS + schema design burden on you	No No concept of agents
Semantic search (embeddings)	No No embedding search	Yes	Yes Archival storage	No Not a vector store	Yes	Yes Core feature	Yes pgvector, Jina v5	Yes Hybrid vector + keyword
Temporal awareness	Yes BEAM architecture	No No temporal layer	Partial Recall storage, limited	No	Yes Conversation timeline	No Timestamps only	No Timestamps only, no reasoning	No Timestamps only
Auto-deduplication	No None built in	Yes Core feature	Partial Agent self-editing	No	Partial Graph dedup	No	No Manual	No
User personalization	No	Partial user_id / agent_id split	No	Yes Dialectic reasoning, cross-app	Partial User session memory	No	No	No
Knowledge graph	No	Partial Pro tier only	No	No	Yes Core feature	No	No	No
Benchmark score (LongMemEval)	91.4% Highest published	49.0% Published	None published	N/A Not a memory system	None published	N/A Not a memory system	N/A Not a memory system	N/A Not a memory system
Lock-in risk	Low MIT, portable	Low OSS core	High Agents live inside Letta	Low Complementary, additive	Medium Cloud-first design	Low Portable embeddings	Low Postgres-compatible	Low Open format
Best for	Highest accuracy memory	Quick bolt-on memory	New agents from scratch	User personalization	Conversation memory + KG	Local dev / prototyping	RAG + SQL workloads	Serverless hybrid search
Watch out for	No embedding search	Broken AND queries	High lock-in, pre-1.0	Not a storage backend	Cloud-first pricing	Not built for prod scale	All plumbing on you	Maturing ecosystem

Data sourced from official docs and published benchmarks — April 2026. Corrections welcome.

Honest Recommendations

Which one should you use?

The right answer depends on your scale, your use case, and what you are actually trying to solve. No single tool wins across all dimensions.

You need the highest published memory accuracy for a single agent or small team

Use

Hindsight

91.4% LongMemEval (highest published). BEAM temporal tracking. MIT licensed. Pair with a vector DB if you need semantic search.

You want automatic memory extraction without building anything from scratch

Use

Mem0

Pluggable, fast setup, handles fact extraction and dedup automatically. Budget for the AND query bug if you need multi-entity filtering.

You need a knowledge graph extracted from conversations

Use

Zep

The only tool here with first-class knowledge graph extraction from conversations. Production-ready.

You are building a consumer product and need persistent user personalization

Use

Honcho

Dialectic reasoning over user behavior, cross-app profiles. Use alongside a storage backend — Honcho is not a database.

You need RAG with full SQL power and complete data portability

Use

Supabase pgvector

Production-grade, Postgres-compatible, Jina v5 embeddings. Best if your team is already in SQL and you want semantic search alongside it.

You are prototyping locally and need the simplest possible setup

Use

ChromaDB

Simplest Python API, local-first, zero infrastructure. Not production scale, but perfect for getting started fast.

You are building a production multi-agent system and need temporal + semantic + dedup + personalization

Use

Combine tools

No single tool covers all dimensions. A common production stack: Hindsight (temporal memory) + Supabase pgvector (semantic search) + Mem0 (dedup) + Honcho (personalization).

Back to the full landscape.

All 13 tools, categories explained, and a quick-compare matrix.

← Back to Memz