--- title: "9 Best Memory Layers for AI Agents in 2026" description: "Nine memory layers for AI agents compared for production in 2026, including past.dev, Hindsight, mem0 and Zep, on changing facts, sources and benchmarks." canonical: https://past.dev/blog/best-memory-layers-for-ai-agents date: 2026-11-16 last-updated: 2026-11-16 category: Engineering authors: The past.dev team --- # 9 Best Memory Layers for AI Agents in 2026 The best memory layer for AI agents returns what is true now, what it replaced and where it came from, inside the stack a team already runs. For production agents that work from changing business records, past.dev ranks first on BEAM at every history size from 100K to 10M tokens among published results ([self-reported run of 29 September 2026, and the harness is public](/benchmarks/beam)). Hindsight, mem0, Zep, Cognee, Graphiti, Letta, Supermemory and LangMem fit narrower setups, each covered below. This best AI agent memory layer frameworks comparison scores nine platforms on five production criteria and ends with when to choose each one. For background on the category, see [What is agent memory?](/what-is-agent-memory) ## How We Compared These Memory Layers Two current trends in AI agent long-term memory solutions shape the criteria. AI agents now work from business records whose facts change, such as emails, tickets and meeting notes. Teams also need answers they can trace back to a source. Every memory layer below is scored on the same five points. - **Changing facts:** whether a later value supersedes an earlier one while the earlier value stays readable as history. - **Source provenance:** whether each result comes back with its date and the verbatim excerpt of the source behind it. - **Benchmark evidence:** whether the vendor publishes results on BEAM, the largest public memory benchmark (ICLR 2026), which asks 2,000 questions over histories of 100K to 10M tokens. - **Framework support:** how the platform connects to the agent frameworks and coding tools a team already uses. - **Self-hosting and deployment:** who runs the system: a managed service, your own infrastructure, or both. Platforms with published BEAM results come first in the list below. ## Top AI Agent Memory Platforms 2026 at a Glance The BEAM column shows each platform's published results at 1M and 10M tokens, the two sizes most vendors report, as of November 10, 2026. Each figure is the vendor's own run, so models, judges and configurations differ between rows; read them as indicative. past.dev's figures are self-reported, and the harness is public. Results at every history size are on the [benchmarks page](/benchmarks). | Platform | How it runs | Licence | Published BEAM results (1M / 10M tokens) | Best for | |---|---|---|---|---| | **past.dev** | Managed service; Bring your own cloud on Enterprise | Commercial engine; open-source plugins | 90.65% / 85.03% (self-reported) | Agents that work from changing business records | | **Hindsight** | Docker image, Python server or embedded library | MIT | 79.1% / 64.1% | Self-hosted memory with published results | | **mem0** | Library, self-hosted server or managed platform | Apache 2.0 | 64.1% / 48.6% | Open-source memory you can self-host | | **Zep** | Managed service built on Graphiti | Commercial service | None published | A managed temporal knowledge graph | | **Cognee** | Engine your team operates | Apache 2.0 | Not compared here | Knowledge graphs built from your documents | | **Graphiti** | Library on a graph database you operate | Apache 2.0 | None published | Temporal facts inside your own stack | | **Letta** | Agent platform with built-in memory | Apache 2.0 | None published | New agents built on one platform | | **Supermemory** | Hosted API | MIT repository | None published | Many file types through connectors | | **LangMem** | Utilities on a LangGraph store you operate | MIT | None published | LangGraph apps with simple memory | ## The 9 Best Memory Layers for AI Agents in 2026 ### 1. past.dev [past.dev](/) is the memory layer for AI agents. An application sends it timestamped text, such as emails, documents, meeting transcripts and support tickets, and past.dev derives memory from that text. When an agent needs context, it recalls as a named identity and gets back the memories that answer its question, ranked, each with its date and the verbatim excerpt of its source. A later value supersedes an earlier one, and the earlier value stays readable as history. - **How it runs:** a managed service you call through the [Memory API](/docs/memory-api/overview), with four core endpoints for ingest, status, deletion and recall. On an Enterprise agreement, it can also run with your own AWS, Google Cloud or Azure account. - **Licence:** commercial engine. The agent plugins, integration packages and benchmark harness are open source. - **Published results:** first on BEAM at every history size among published results: 92.08% at 100K tokens, 89.63% at 500K, 90.65% at 1M and 85.03% at 10M. The results are self-reported (run of 29 September 2026) and the harness is public. - **Strengths:** current and previous values with dates and sources; recall scoped to the identity asking, with audiences that set what each reader may see; an MCP server and plugins for Claude Code, Codex, Cursor, JetBrains IDEs (Junie) and Hermes Agent; SOC 2 Type II audited, ISO 27001 and ISO 27701 certified. - **Limits:** self-hosting is not offered, and Bring your own cloud requires an Enterprise agreement. Its lowest BEAM scores are in multi-session reasoning. - **Best for:** agents that work from changing business records and need the current value, its source and per-user access control. ### 2. Hindsight [Hindsight](https://hindsight.vectorize.io/), from Vectorize, is an MIT-licensed memory system built on three operations: retain, recall and reflect. Reflect generates a response from stored memory. - **How it runs:** as a Docker image, a Python server or an embedded library, on PostgreSQL with pgvector or on Oracle. - **Licence:** MIT. - **Published results:** BEAM 86.2% at 100K tokens, 80.1% at 500K, 79.1% at 1M and 64.1% at 10M, as published on agentmemorybenchmark.ai. - **Strengths:** the second-highest published BEAM results at 100K and 500K tokens, and several deployment shapes. - **Limits:** your team operates the server and the database. - **Best for:** teams that want a self-hosted, open-source memory layer with published benchmark results. ### 3. mem0 [mem0](https://docs.mem0.ai/) is an Apache 2.0 memory layer with Python and TypeScript SDKs. Applications add memories and search them, with memory scoped by user, session and agent and stored in a vector store you configure. - **How it runs:** as a library, a self-hosted server or a managed platform. - **Licence:** Apache 2.0. - **Published results:** BEAM 64.1% at 1M tokens and 48.6% at 10M tokens, from mem0's own runs on its managed platform. No figure at 100K or 500K. - **Strengths:** SDKs in two languages and three ways to deploy. - **Limits:** self-hosting means operating the server and the vector store. - **Best for:** teams that want an open-source memory library they can self-host today. ### 4. Zep [Zep](https://help.getzep.com/) is a managed memory layer for AI agents built on Graphiti. It stores what an agent learns as facts with validity windows and links each fact to the source episode it came from. - **How it runs:** as a managed memory service; the graph runs on Neo4j, FalkorDB or Amazon Neptune. - **Licence:** commercial service. The Graphiti engine underneath is Apache 2.0. - **Published results:** none. We found no BEAM result published by Zep as of November 10, 2026. - **Strengths:** agent memory with time recorded on each fact and links back to the source episodes. - **Limits:** no published BEAM result to compare it with other memory layers. - **Best for:** teams that want a managed temporal knowledge graph as the memory layer for their AI agents. ### 5. Cognee [Cognee](https://docs.cognee.ai/getting-started/introduction) is an open-source memory engine for AI agents. It builds a knowledge graph with embeddings from the data you supply, and agents query that graph as their memory. - **How it runs:** in infrastructure your team operates. - **Licence:** Apache 2.0. - **Published results:** not compared here. - **Strengths:** turns your own documents into a knowledge graph your agents can query as memory. - **Limits:** your team runs and operates the memory engine. - **Best for:** teams that want agent memory built from their own documents and have the capacity to run it. ### 6. Graphiti [Graphiti](https://github.com/getzep/graphiti) is an Apache 2.0 temporal knowledge graph framework from the Zep team that developers use to build their own memory layer for AI agents. It is a library that runs on a graph database you operate and needs a model provider key. - **How it runs:** as a memory library inside your application, on your own graph database. - **Licence:** Apache 2.0. - **Published results:** none found as of November 10, 2026. - **Strengths:** time recorded on facts, open-source code and the same engine that underpins Zep. - **Limits:** your team operates the graph database and pays for the model calls. - **Best for:** teams with graph infrastructure that want temporal agent memory inside their own stack. ### 7. Letta [Letta](https://docs.letta.com/guides/agents/memory-blocks), formerly MemGPT, is an Apache 2.0 platform for stateful agents. It keeps memory blocks in the agent's context plus archival memory that the agent searches through tools, and the agent writes its own memory. - **How it runs:** as a full agent platform. - **Licence:** Apache 2.0. - **Published results:** none published. - **Strengths:** memory and agent runtime in one platform, with the agent managing its own memory. - **Limits:** adopting Letta's memory means adopting its runtime. - **Best for:** teams building a new agent that want runtime and memory together. ### 8. Supermemory [Supermemory](https://supermemory.ai/docs/intro) is context infrastructure for agents. It ingests text, chats, files, PDFs, images and video through extractors and connectors and serves them through a hosted API. - **How it runs:** as a hosted API. - **Licence:** MIT repository. - **Published results:** no BEAM figure published. - **Strengths:** broad input types and ready-made connectors. - **Limits:** no BEAM result to compare against. - **Best for:** apps that need to ingest many file types through connectors. ### 9. LangMem [LangMem](https://langchain-ai.github.io/langmem/) provides utilities that extract memories from conversations into a LangGraph store the application operates. Entries are namespaced, and writing a key replaces it. - **How it runs:** inside a LangGraph application, on a store you operate. - **Licence:** MIT. - **Published results:** none published. - **Strengths:** a close fit for apps already built on LangGraph. - **Limits:** writing a key replaces the earlier value, so the earlier value is not kept as history. - **Best for:** LangGraph apps that need simple memory inside the framework. ## How to Choose a Memory Layer for Production Choosing the best memory layer for AI agents in production comes down to who operates it and what the agent must remember. Several options combine: a Letta tool or a LangGraph app can call past.dev for recall. The [guide to choosing a memory system](/guides/choose-memory-system) covers the decision in more depth. - **Choose past.dev if** your agent works from changing records, needs the current value with its dated source, and must limit what each user sees, with no infrastructure to run. - **Choose Hindsight if** you want an MIT system you run yourself, with published BEAM results. - **Choose mem0 if** you want an Apache 2.0 library you can self-host today, with a managed option later. - **Choose Zep if** you want a managed temporal knowledge graph. - **Choose Cognee if** you want a knowledge graph built from your documents and you are ready to operate the engine. - **Choose Graphiti if** you already operate a graph database and want temporal facts as a library. - **Choose Letta if** you are building a new agent and want the runtime and memory on one platform. - **Choose Supermemory if** you need to ingest many file types through ready-made connectors. - **Choose LangMem if** your agent runs on LangGraph and per-key memory is enough. ## Sources for This Comparison of Memory Layers for AI Agents Each memory layer's details in this best AI agent memory frameworks comparison come from the vendor's own documentation and repositories as of November 10, 2026, and describe how it stores, updates and recalls what an agent needs to remember. Agent memory layers change fast, so check each vendor's current docs before choosing one. ## FAQs About Memory Layers for AI Agents ### What's the Best Memory Layer for AI Agents? For production agents that work from changing business records, past.dev is the strongest fit. It returns current and previous values with dated sources, scoped to the identity asking, and ranks first on BEAM at every history size among published results (self-reported, with a public harness). Teams that need open-source software they host themselves should look at Hindsight or mem0. ### What Are the Top Agent Memory Platforms Right Now? The top agent memory platforms 2026 covered here are past.dev, Hindsight, mem0, Zep, Cognee, Graphiti, Letta, Supermemory and LangMem. They fall into three groups: managed services (past.dev, Zep, Supermemory), open-source systems you can run yourself (Hindsight, mem0, Cognee, Graphiti) and framework-level tools (Letta, LangMem). ### Which Agent Memory Frameworks Are Worth Using in Production? Three questions decide the shortlist: do facts keep their history, do answers carry sources, and are benchmark results published. Among the platforms in this comparison, past.dev (self-reported, with a public harness), Hindsight and mem0 have BEAM results shown on past.dev's benchmarks page. For the others, test against your own data before committing. ### Which Agent Memory Projects Are Most Popular on GitHub? Open-source projects such as mem0, Graphiti and Hindsight draw wide attention on GitHub. Star counts measure interest, so pair them with benchmark results and a check of how each memory layer handles changing facts. past.dev's engine is commercial; its plugins and benchmark harness are open source on GitHub. ### How Is a Memory Layer Different from RAG? RAG retrieves text that looks like the question and puts it in the prompt. A temporal memory layer such as past.dev returns what is true now, what it replaced and the source for each, so the agent can tell a current fact from an outdated one. Applications often combine a vector database with a memory layer. ## Try past.dev To test past.dev as the memory layer for your AI agents, [get an API key](https://sso.past.dev/sign-up). Sign up and start with a one-time grant of credits, more with a work email. No card required, and nobody approves the account.