Zep
Zep is an agent memory platform for developers: it turns chat, business data and documents into a temporal Context Graph and returns a prompt-ready context block in under 200 ms, through Python, TypeScript and Go SDKs.
What is Zep?
Zep is memory infrastructure for AI agents, published by Zep Software, Inc. in San Francisco. Every piece of data sent to it, whether a conversation turn, a business record, a document or a JSON payload, is treated as an Episode. Zep extracts entities and facts from that episode and files them into a temporal Context Graph built by its proprietary Context Graph Engine, in which isolation and time are native properties of the data model rather than layers added afterwards.
Time is the distinguishing idea. When new information contradicts what the graph already holds, the older fact is invalidated but kept as history, so an agent can ask what is true now or what was true on a given date. Every fact carries provenance back to the episode that produced it, and an Observations layer reads the structure of the graph to surface recurring patterns and co-occurrences that no single fact states outright.
Above the individual graph sits the Context Lake: millions of context graphs, one per user, customer, team or topic, governed and served as a single system, the data-lake pattern applied to agent context. Smart Context Assembly then selects facts, summaries and observations against a token budget and returns a prompt-ready block. Zep publishes p95 retrieval under 200 ms regardless of graph size, from 148 ms at 10,000 entities to 168 ms at 100 million, plus benchmark results of 94.7% on LoCoMo and 90.2% on LongMemEval with the methodology online.
Integration is a developer job: SDKs for Python, TypeScript and Go, a REST API, integration packages for 14 agent frameworks, and a Memory MCP Server for using the same memory from Claude, ChatGPT or Cursor. The underlying temporal-graph framework, Graphiti, is open source. Governance is built in, with RBAC for people, ABAC for agents and MCP users, retention policies, audit and API logs and enterprise SSO, and three deployment models are offered: managed Cloud, Cloud with customer-held AWS KMS keys (BYOK), and BYOC inside the customer's own VPC.
The company was founded in 2023, went through Y Combinator's W24 batch, and cites an April 2026 S&P Global Market Intelligence report on agent memory. The stated limits are equally clear: this is an API product with no no-code interface, and the strongest compliance guarantees sit on the Enterprise plan.
What it does
- Ingest any data source as an Episode: chat message, JSON payload, block of text or document.
- Build a temporal knowledge graph automatically, one per user, customer, team or topic.
- Return a ranked, token-budgeted context block ready to drop into a prompt.
- Invalidate facts contradicted by newer information while keeping the superseded version as history.
- Query the graph as it stands now, or as it stood on a past date.
- Trace every fact back to the source episode, and surface Observations, the recurring patterns and co-occurrences found in the graph.
- Govern access with ABAC policies, retention schedules with legal hold, audit logs and API logs.
When to use Zep / When not to
A quick filter to help you decide if Zep is the right fit.
When to use Zep
- Engineering teams shipping AI agents to production that need state to persist across sessions, and that must combine conversation history with business data in a single queryable graph.
- Regulated organizations that require SOC 2 Type II, a HIPAA Business Associate Agreement or a BYOC deployment inside their own VPC.
- Teams already standardized on an agent framework: 14 frameworks, including LangGraph, CrewAI, AutoGen, Google ADK, Pydantic AI and the Vercel AI SDK, ship with an integration package.
- High-volume products where retrieval latency must stay predictable, since Zep publishes p95 retrieval of 148 ms at 10,000 entities and 168 ms at 100 million.
- Solo developers and early-stage startups prototyping agent memory on the permanent free tier of 10,000 credits per month.
When not to use Zep
- Non-technical users looking for a no-code memory tool: everything goes through the SDKs or the REST API, so writing code is unavoidable.
- Anyone expecting a mobile, desktop or browser-extension product, since Zep ships an API and a web dashboard only.
- Buyers who require EU data residency on a managed service: every listed subprocessor processes in the United States and no EU region is offered.
- Organizations that must contractually rule out any use of their data for model training, because Terms 5.2 grant Zep that licence and no opt-out is documented.
- Teams that need SOC 2 Type II evidence, a HIPAA BAA, audit logs or a DPA on a self-serve budget, as all four sit on the Enterprise plan.
How to use Zep
A typical end-to-end flow, from setup to results.
- Create an account on the Zep dashboard; sign-up opens the free tier of 10,000 credits per month.
- Install the Python, TypeScript or Go SDK, or call the REST API directly.
- Try the flow in the dashboard Playground before writing application code.
- Run the canonical loop: create a user, create a thread, add messages, then get the context.
- Push business data into the same user's graph with graph.add and a typed JSON payload.
- Retrieve the prompt-ready block with thread.get_user_context and pass it to your model.
- If you use one of the supported agent frameworks, install its integration package instead; it wires Zep into the framework's native memory patterns.
- Alternatively, skip application code entirely and connect the Memory MCP Server to Claude, ChatGPT or Cursor.
- Tune extraction with custom entity and edge types, custom extraction instructions and webhooks as your schema stabilizes.
- Lean on the quick-start guide, the cookbooks, the zepctl CLI and debug mode while iterating.
Pros & Cons
Pros
- Published p95 retrieval under 200 ms that stays flat from 10,000 to 100 million entities, moving only from 148 ms to 168 ms.
- Public benchmark results with the methodology online: 94.7% on LoCoMo and 90.2% on LongMemEval, with a more compact context than the alternatives cited, at 5,760 and 4,408 tokens.
- Governance built into the substrate rather than bolted on: ABAC, retention policies, audit trails and per-fact provenance.
- Three deployment models, including BYOC inside the customer's own VPC, plus BYOK encryption with customer-held AWS KMS keys.
- SOC 2 Type II certification and a HIPAA BAA available, alongside a named and dated subprocessor annex giving purpose and country of processing for each entry.
- Integration packages for 14 agent frameworks, and an open-source engine, Graphiti, that can be inspected independently.
- Predictable entry point: a permanent free tier of 10,000 credits per month, with retrieval, storage, threads, users and graph storage unmetered on every plan.
Cons
- Terms 5.2 grant Zep a perpetual, irrevocable licence over Customer Data that expressly includes training models, and no opt-out mechanism is documented.
- No EU hosting region on the managed service: all 20 listed subprocessors process in the United States.
- SOC 2 Type II evidence, the HIPAA BAA, audit logs and the DPA are Enterprise-plan items, out of reach on the self-serve tiers.
- The Trust Center is an empty JavaScript shell, so the certifications it is meant to evidence cannot be verified online.
- The privacy policy is still version 1.0 dated 27 January 2024, and the word GDPR appears nowhere on the site.
- Credit pricing scales with Episode size, so the real monthly cost has to be estimated before committing.
- Community support is the only option included on the first paid plan, and neither contact nor demo has a page of its own since both are modals.
Pricing & Plans
Zep operates a credit-based freemium model. A permanent free tier provides 10,000 credits per month, with 2 projects, 1 Memory MCP Server seat, no rollover and no automatic top-up; the same 10,000 monthly credits also appear as free trial credits on the paid plans, as a recurring monthly allowance rather than a time-limited trial. The lowest paid entry point is the Flex plan at USD 125.00 per month billed monthly, or USD 1,250 per year, that is USD 104 per month, with the annual commitment advertised as 17% cheaper. Flex Plus is USD 375.00 per month, or USD 3,750 per year, that is USD 312 per month. Credits are consumed on ingestion and processing only, at one credit per Episode of up to 350 bytes plus one credit per additional 350-byte block started, and one eighth of a credit per webhook call, while retrieval, storage, threads, users and graph storage are unmetered. Overage is sold at USD 25 per 10,000 credits on Flex and USD 75 per 40,000 credits on Flex Plus, with automatic top-up below a 20% balance and credit rollover of 30 days on Flex and 60 days on Flex Plus. Enterprise pricing is on quotation, with negotiated credits, negotiated rates and an SLA. As a worked example, the site's estimator prices 15,000 Episodes of 700 bytes plus 20,000 webhook calls per month at 32,500 credits, or roughly USD 125 per month on Flex.
- 10
- 000 credits per month
- 2 projects
- 1 Memory MCP Server seat
- 5 custom entity and edge types
- rate limits that vary with service load
- low-priority Episode processing
- no rollover and no automatic top-up.
- USD 1
- 250 per year (USD 104 per month billed annually) or USD 125 per month
- for 50
- 000 credits per month
- 600 requests per minute
- 5 projects
- 5 MCP seats
- 10 custom types
- USD 3
- 750 per year (USD 312 per month billed annually) or USD 375 per month
- for 200
- 000 credits per month
- 1
- 000 requests per minute
- 10 projects
- 15 MCP seats
- on quotation
- with negotiated credits and rates
- guaranteed rate limits under an SLA
- unlimited projects
- custom MCP seats
- SOC 2 Type II
- HIPAA BAA
- audit logs
- unlimited memories
- retrieval and users
- with the temporal Context Graph included at every tier.
Data, GDPR & hosting
A consolidated view of how Zep handles your data.
GDPR overview
The word GDPR does not appear anywhere on Zep's website: zero occurrences across the pages collected, and the company never claims compliance in writing. The closest statement sits in the pricing FAQ: "Yes. Zep signs Data Processing Agreements with EU customers and customers with EU presence." Terms section 7 confirms that a signed DPA is incorporated into the agreement and prevails on personal-data processing. A dated subprocessor annex, updated 17 August 2026, lists each subprocessor, its purpose and its country of processing. Against that: the DPA sits on the Enterprise line of the plan table, no Article 27 EU representative is designated, no DPO or GDPR contact address is published since the privacy policy routes all questions to the general information mailbox, and that policy is still version 1.0 of 27 January 2024, detailing only marketing opt-out rather than the wider set of data-subject rights.
Who owns the data?
Under Terms section 5.2, the customer retains all right, title and interest in its Customer Data, defined in 1.5 as Inputs, Outputs and any other content submitted, including through third-party services. In the same clause the customer grants Zep Software, Inc. a non-exclusive, worldwide, perpetual, irrevocable, royalty-free and fully paid-up licence to access, use, reproduce, modify and create derivative works from that data for any lawful purpose, expressly including training and improving machine-learning and artificial-intelligence models. Two derived categories belong to Zep outright: Aggregate Data, aggregated and de-identified, under 5.3, and Performance Data, technical and usage logs, under 5.4. Section 10.3 states that both of those rights survive termination.
Reuse rights
Because ownership stays with the customer under Terms 5.2, the customer may reuse, export and redistribute its own Inputs and Outputs without asking Zep for permission. On Zep's side, the licence granted in the same clause covers any lawful purpose, including operating the service and training and improving machine-learning and artificial-intelligence models; no mechanism for opting out of that training licence is documented on the site or in the documentation. Terms 2.7 add that Inputs are shared with the third-party services supplying the AI Tools used to generate Outputs, and the subprocessor annex names OpenAI L.L.C., Google LLC for Gemini, Thinking Machines Lab Inc. and Deep Infra, Inc. for that processing. Terms 2.6 further state that Zep and those providers are not each other's processors under applicable privacy laws. Bring Your Own Model lets a customer route inference through its own OpenAI, Anthropic, Google, AWS Bedrock or Azure credentials, which changes who performs inference but leaves the training licence Zep holds unchanged. Separately, the privacy policy states that personal data collected through the website is neither sold nor rented for marketing, and may be shared with service providers acting on Zep's instructions.
Data retention & training
Hosting summary
Zep publishes a subprocessor annex dated 17 August 2026 listing 20 named subprocessors, every one of them with United States as the country of processing. Amazon Web Services, Inc. supplies the cloud infrastructure "used to host and operate Zep's platform, including storage of customer data and execution of application workloads", and Microsoft Corporation, through Azure, provides compute and storage for model inference and data processing. LLM processing involves OpenAI L.L.C., Google LLC for Gemini, Thinking Machines Lab Inc. and Deep Infra, Inc. The remaining subprocessors cover task orchestration, documentation assistance, authentication, product analytics, observability, payments, usage billing, webhooks, email delivery, CDN and WAF, with Hatchet Technologies, Kinde, PostHog, Datadog, Stripe, Schematic, Svix, Loops, Resend, Cloudflare, Microsoft Clarity and LinkedIn Corporation among them. No EU or multi-region hosting option is offered on the managed service. Two alternatives exist: BYOC, which runs Zep inside the customer's own VPC, what the documentation FAQ calls VPC residency, and BYOK, which encrypts data at rest with a customer-held, revocable AWS KMS key. The domain's IP resolves to a Cloudflare anycast node and therefore says nothing about where data actually sits.
Things to keep in mind
Risks and trade-offs to weigh before adopting Zep.
- Terms 5.2 grant a perpetual, irrevocable licence over Customer Data that includes model training, with no published opt-out: whatever your agent stores about your end users falls inside that grant.
- Aggregate Data under 5.3 and Performance Data under 5.4 belong to Zep outright, and section 10.3 states those rights survive termination, so leaving the service does not undo them.
- Inputs are passed to third-party AI Tool providers under Terms 2.7, while 2.6 states that Zep and those providers are not each other's processors, which complicates the accountability chain for personal data.
- No hosting country other than the United States is offered on the managed service, and no EU representative is designated, so European deployments rely on a DPA or a BYOC installation.
- A persistent memory layer accumulates a durable profile of each end user across sessions, and no fixed retention period is published for Customer Data, so decide deliberately what is worth remembering and set retention policies rather than defaulting to keeping everything.
- The Trust Center presented as the source of compliance attestations returns no readable content, the privacy policy has not been revised since 27 January 2024, and the corporate name differs between the 2024 legal documents (Zep Software, Inc.) and the 2026 footer (Zep AI, Inc.): confirm certifications and contracting entity directly with the vendor.
- Credit cost follows Episode size, since a 1,200-byte Episode costs 4 credits, and free-tier and Flex rate limits can be adjusted by Zep according to service load, so plan for cost drift and for throughput that is not guaranteed below Enterprise.
Setup & Integrations
Technical difficulty
Developer-level: there is no no-code interface, everything runs through the Python, TypeScript or Go SDK or the REST API. The core loop is short, being create user, create thread, add messages, get context, and Zep advertises "Three Lines of Code". A quick-start guide, cookbooks, the zepctl CLI and debug mode help; on one of the 14 supported agent frameworks an integration package does the wiring, and connecting the Memory MCP Server to Claude, ChatGPT or Cursor needs no application code. Custom entity types, extraction instructions, webhooks, ABAC policies and BYOC in your own VPC require real engineering.
Deployment
Integrations
Behind Zep
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement Zep.
Frequently asked questions
What is an Episode in Zep?
How are credits consumed?
Is there a free plan?
Do I pay for storage?
Is Zep SOC 2 Type II certified, and can it sign a HIPAA BAA?
Where is the data hosted?
Is my data used to train models?
Does Zep sign a DPA with EU customers?
Which languages and frameworks are supported?
Can I self-host, and is there a mobile app?
Should you pick Zep?
Zep is a technically mature piece of infrastructure. The performance claims are specific and testable, with p95 retrieval under 200 ms from 10,000 to 100 million entities, 94.7% on LoCoMo and 90.2% on LongMemEval, the benchmark methodology is published, and the engine behind the memory, Graphiti, is open source. Governance is part of the data model rather than an afterthought: per-fact provenance, ABAC policies, retention schedules with legal hold, audit logs and API logs. For a team already running agents in production, that combination is unusual.
The commercial terms read clearly at the self-serve tiers and less clearly where it counts, because credits are consumed by Episode size, so budgeting means estimating the average payload before committing.
Three points deserve a decision before adoption rather than after. First, Terms 5.2 grant Zep a perpetual, irrevocable, royalty-free licence over Customer Data that expressly includes training and improving machine-learning models, with no documented opt-out; this is the right the contract confers, not an observed practice, but it is the right a signature accepts. Second, there is no EU hosting option on the managed service, since all 20 listed subprocessors process in the United States, leaving European buyers with a DPA, a BYOC deployment in their own VPC, or neither. Third, the privacy documents date from January 2024, the word GDPR appears nowhere on the site, and the Trust Center meant to hold the attestations returns nothing readable.
Finally, the strongest assurances, namely SOC 2 Type II, the HIPAA BAA, audit logs and the DPA, sit behind the Enterprise plan, which is on quotation. Zep suits engineering teams that can accept US processing and negotiate their contract; teams with strict data-residency or training-exclusion requirements should settle those clauses first.
- Choosing a selection results in a full page refresh.
- Opens in a new window.