Which AI memory tool should an agent team start with?
Start with the smallest system that can pass your own memory evaluation. A managed API is the shortest path when the team needs add, extract, search, scope and delete operations without operating a vector database and graph pipeline. A framework such as Letta or a graph toolkit such as Cognee is the better test when the agent runtime itself is being redesigned or the ontology is the product.
This comparison is for application teams building assistants, coding agents, support agents and voice agents. It evaluates the memory layer separately from the LLM, application framework and vector database. Prices and features were checked against public vendor pages on October 20, 2026 and should be rechecked before purchase.
How do the leading AI memory tools compare?
Mem0, Zep and Supermemory lead with a managed memory API. Letta combines an agent runtime with editable memory blocks. Cognee emphasizes a graph and ontology pipeline that can be self-hosted. LangMem provides building blocks for long-term memory inside LangGraph-oriented applications. AI Memory focuses on a model-neutral memory layer with event ingestion, extraction, merge, compression, graph relations, scoped retrieval and export.
| Tool | Primary shape | Best fit |
|---|---|---|
| AI Memory | Managed memory API; graph, merge, compression and scoped retrieval | Teams that want durable user and agent memory without adopting a new agent runtime |
| Mem0 | Managed or open-source memory layer with vector and graph options | Teams that want a broad ecosystem, many integrations and an open-source path |
| Zep | Managed context graph with temporal reasoning and Graphiti tooling | Teams prioritizing bi-temporal facts and graph-first context assembly |
| Letta | Stateful agent framework with editable memory blocks | Teams willing to adopt an agent operating model and runtime |
| Cognee | Self-hostable graph and vector pipeline with ontology-oriented tools | Teams that need to assemble and govern a domain knowledge graph |
| Supermemory | Managed memory API, profiles, graph memory and connectors | Teams that want a broad personalization API and connector surface |
| LangMem | Open-source memory utilities for LangGraph applications | Teams already standardized on LangGraph and prepared to operate the stack |
Product classifications are simplified. Verify current capabilities and limits in each vendor’s official documentation.
What do AI memory tools cost?
The entry price is only the start of the comparison. Memory products meter different units: stored memories, add requests, retrieval requests, episodes, credits, tokens or a dollar balance. Zep’s public pricing uses episode-size credits, Mem0 separates add and retrieval requests, and Supermemory prices plans with a monthly usage balance. AI Memory’s public plans use stored-memory limits rather than charging separately for every retrieval.
Normalize each vendor against one workload: 10,000 active users, 20 new events per user per month, three retrievals per active session and a defined retention period. Ask what happens to storage, graph maintenance, re-embedding and model calls when an event grows from 300 bytes to 3,000 bytes.
- Mem0’s public Hobby tier is free with request limits; paid tiers are documented on its pricing page.
- Zep’s self-serve plans use credits; episode size changes credit consumption.
- Supermemory publishes Free, Pro, Max and Scale plans with a monthly usage balance.
- Letta and Cognee include self-hostable paths, but infrastructure and operations remain your cost.
- AI Memory publishes Pro and Scale plans with monthly or annual billing through Stripe.
How should a team evaluate memory quality?
Build a 30–50 conversation evaluation set from your own product. Include a changed preference, a superseded fact, two users with similar names, a tenant boundary, a request to forget and a question whose answer needs two related memories. Score direct recall, irrelevant retrieval, contradiction handling, answer quality after retrieval, latency and tokens passed to the model.
The vendor should expose the records it retrieved, not only the final answer. Without that visibility, a better LLM response can hide a memory system that recalled the wrong user or ignored a deletion. A useful demo proves the full loop: write, extract, retrieve, correct, and delete.
- Fact recall: did the correct durable fact reach the prompt?
- Precision: how many retrieved memories were irrelevant?
- Temporal truth: did the system prefer the newer fact without losing history?
- Isolation: could a query cross a user, agent or tenant boundary?
- Forgetting: did deletion remove the fact from future retrieval and export?
- Portability: can the team retrieve a complete, useful export?
Which memory tool has the strongest governance story?
Governance is a set of observable controls, not a security badge. Require query-time scope enforcement, per-user deletion, tenant deletion, retention rules, export, audit history and a documented data path. A graph platform may offer stronger provenance while a managed API may reduce operational exposure; neither replaces an application-level data map.
AI Memory treats correction and deletion as first-class operations. Cognee’s public material emphasizes self-hosting and regulated deployments, Zep emphasizes governed context graphs, and Mem0 exposes project and organization controls. Match the control plane to your actual compliance obligations and test it with the same rigor as retrieval.
What is the shortest defensible selection process?
Shortlist two products, run the same private evaluation set and calculate monthly cost at your expected usage. Choose the product that passes recall and isolation with the least operational work. Sign a small production plan only after export, correction and deletion are proven.
If the agent runtime is already stable, adding a model-neutral memory layer preserves your architecture. If the agent itself needs a new stateful runtime, evaluate Letta. If ontology and graph construction are the core project, start with Cognee. If temporal graph reasoning is the priority, evaluate Zep. If broad managed integrations matter most, compare Mem0, Supermemory and AI Memory.
Frequently asked questions
What is the best AI memory tool overall?
No single product is best for every agent. Choose a managed API for faster integration, a graph platform for ontology-heavy work, or an agent framework when the runtime itself is changing. Evaluate the candidates on your own conversation set.
Is a vector database an AI memory layer?
A vector database stores and retrieves embeddings, but a memory layer also handles extraction, merging, contradictions, scopes, retention, correction and deletion. Many memory products use a vector database as one component.
Should I self-host agent memory?
Self-host when data residency, network boundaries or customization justify the operational cost. Use a managed layer when the team would rather spend engineering time on the product than on storage, upgrades, backups and retrieval infrastructure.
How do I avoid vendor lock-in?
Require a complete export of durable facts, relations and metadata; test import into a local store; and verify deletion and retention controls. The ability to leave is part of the product architecture.
How often should this comparison be updated?
Quarterly, or whenever a vendor changes pricing, deployment options or core capabilities. Memory products are moving quickly, so an undated comparison becomes misleading fast.