Zep Review 2026: Pricing, Features, Pros & Cons
Zep is a managed memory layer for AI agents built on temporal context graphs rather than a vector store with similarity search bolted on. The architecture addresses a real and under-discussed failure — stale facts winning over current ones — and the compliance story at the top end is genuinely enterprise-ready. The obstacle is the entry price: $125 a month, with no free tier to prove the case on first.
Quick Verdict
Best for: long-running agents where user facts change over time and surfacing a stale one is costly, plus regulated buyers who need SOC 2, a HIPAA BAA and VPC deployment. Not for: prototypes, side projects, or anyone who needs to prove the architecture before spending — start with Graphiti instead.
What Is Zep?
Zep is a memory layer for AI agents whose central design decision is temporal. Rather than storing memories as embeddings and retrieving whatever is most semantically similar, it maintains a context graph that tracks facts across time. When a fact is superseded, it is recorded as superseded rather than left to compete with its replacement on similarity score alone.
That distinction matters more than it sounds. A vector store asked “where does this user work” will happily return a statement that was accurate in February and has been wrong since July, because both statements are equally relevant to the question. For an agent that runs for months against the same user, this is the memory failure that actually shows up in support tickets.
Around the graph sit the Context Lake and the Context Graph Engine, with sub-200ms retrieval as the quoted latency budget — the threshold that determines whether memory can live in the request path. The open-source foundation, Graphiti, crossed 20,000 GitHub stars in under a year, so the approach can be inspected rather than taken on faith, and the company publishes results on LoCoMo and LongMemEval instead of asserting quality. Compliance runs to SOC 2 Type II with a HIPAA BAA available and VPC deployment at the Enterprise tier.
Zep Pros & Cons
✓ Pros
- •The temporal context graph solves a specific and expensive failure: an agent retrieving the most semantically similar past message will happily surface a fact that was true six months ago, and Zep records superseded facts as superseded rather than letting them compete with the current one
- •Sub-200ms retrieval is quoted as the latency budget, which is the number that decides whether a memory layer can sit inside the request path of a conversational agent at all rather than running as a background job
- •Graphiti, the open-source foundation, crossed 20,000 GitHub stars in under twelve months — the approach has a public implementation you can read instead of a marketing claim you have to trust
- •The company publishes benchmark results on LoCoMo and LongMemEval, the two standard long-running-agent memory benchmarks, rather than asserting quality in the abstract
- •It ingests chat history, business data and user behaviour into the same graph structure, so memory is not limited to what was said in a conversation
- •Compliance is genuinely enterprise-grade at the top tier: SOC 2 Type II, a HIPAA BAA, and deployment into your own VPC rather than Zep's cloud
- •Credits roll over for 30 days with auto top-up at 20%, which smooths spiky months instead of forcing you to buy for your peak and waste the rest
- •Venture-funded startups can request Enterprise terms at emerging-company pricing, which is a real path to SOC 2 and VPC deployment without an enterprise budget
✗ Cons
- •There is no free tier. The entry price is $125 per month, which means you cannot build a prototype on Zep and decide later whether it was the right architecture
- •Flex is $125/month for 50,000 credits, and additional credits are $25 per 10,000 — so overage costs half again as much per credit as the base allowance, which punishes exactly the growth you are hoping for
- •Credits are an abstraction over the operations you actually care about. You cannot look at the pricing page and directly compute what a thousand daily conversations will cost without modelling the credit consumption first
- •Flex caps you at 5 projects and 10 custom entity and edge types. Custom entity types are the mechanism for making the graph fit your domain, and ten is restrictive for anything beyond a narrow use case
- •API logs are retained for one day on Flex. Debugging a memory issue reported by a user on Friday afternoon is not possible on Monday
- •The jump to Flex Plus is $375/month for 200,000 credits — a three-times price rise for four times the credits, which is better per-unit but a hard step for a growing product
- •Guaranteed rate limits with an SLA are Enterprise-only. Flex gives you 600 requests per minute with no commitment behind it, which is a risk for anything user-facing at scale
- •The open-source Graphiti core is free but is the graph engine, not the managed Context Lake, Context Graph Engine or the operational surface around them — self-hosting is a real project, not a config flag
Zep Pricing 2026
Graphiti (OSS)
- •Open-source temporal graph core
- •20k+ GitHub stars
- •Self-hosted, self-operated
- •No managed Context Lake
- •No SLA or support
Teams who want the architecture without the service
Flex
- •50,000 credits/month
- •600 requests per minute
- •5 projects, 10 custom entity/edge types
- •30-day credit rollover, auto top-up
- •1 day of API logs
First production deployment
Flex Plus
- •200,000 credits/month
- •Same rollover and top-up behaviour
- •Unlimited memories, retrieval and users
- •Extra credits at $25/10,000
- •$312/mo billed annually
Apps past the first growth step
Enterprise
- •Negotiated credit rates
- •Guaranteed rate limits with SLA
- •SOC 2 Type II and HIPAA BAA
- •1-year audit and API log retention
- •VPC deployment, dedicated account manager
Regulated and high-volume production
Plan contents as published on Zep's pricing page in August 2026 — verify before purchase.
The Numbers Behind the Tiers
| Item | Value | What to know |
|---|---|---|
| Flex, billed annually | $1,250/yr ($104/mo) | A 17% saving against monthly billing. Worth taking only once you have proven the architecture, since there is no free tier on which to have proven it |
| Flex Plus, billed annually | $3,750/yr ($312/mo) | Same 17% structure. The per-credit economics improve over Flex — four times the credits for three times the price |
| Additional credits | $25 per 10,000 | Above the base allowance. At Flex's $125 for 50,000, base credits work out at $25 per 10,000 too — so overage is priced at parity rather than at a penalty, which is fairer than it first reads |
| Credit rollover | 30 days | Unused credits survive a month, with auto top-up triggering at 20% remaining. Smooths spiky usage instead of forcing you to provision for peak |
| Retrieval latency | Sub-200ms (quoted) | The number that determines whether memory sits in the request path or has to run asynchronously. Verify it against your own payload shape during evaluation |
Credits are the metering unit for both ingestion and retrieval. Model consumption against real traffic for a fortnight before committing annually.
Who Should Actually Use Zep
Use it if: your agent runs long enough that user facts change and a stale memory would produce a confidently wrong answer; you need memory retrieval inside the request path rather than as a background job, where sub-200ms is the operative constraint; you are a regulated buyer needing SOC 2 Type II, a HIPAA BAA or VPC deployment; or you are venture-funded and can access Enterprise terms at emerging-company pricing.
Skip it if: you are still validating whether your agent needs persistent memory at all — $125 is a lot to spend on a question a free tier elsewhere would answer; if your sessions are short enough that facts never go stale, in which case the temporal machinery is complexity you pay for and do not use; if one day of API log retention would leave you blind when debugging, and Enterprise is out of reach; or if you have the engineering capacity to run Graphiti yourself and would rather own the stack.
Frequently Asked Questions
How much does Zep cost in 2026?
Flex is $125 per month billed monthly, or $104 per month billed annually at $1,250 per year, a 17% saving. It includes 50,000 credits per month, additional credits at $25 per 10,000, auto top-up at 20%, 30-day credit rollover, 600 requests per minute, 5 projects, 10 custom entity and edge types, 1 day of API log retention, and unlimited memories, retrieval and users. Flex Plus is $375 per month billed monthly, or $312 per month billed annually at $3,750 per year, and includes 200,000 credits. Enterprise is custom-priced with negotiated credit rates, guaranteed rate limits with an SLA, unlimited projects, SOC 2 Type II, a HIPAA BAA, one year of audit and API log retention, Slack or Teams support and a dedicated account manager, deployed either in Zep's cloud or your own VPC. Venture-funded startups can request Enterprise at emerging-company pricing. The Graphiti open-source project is free.
What is a temporal knowledge graph and why does it matter for agent memory?
Most memory layers are a vector store with similarity search on top: you ask for context, and you get back the past messages most semantically similar to the current one. The failure mode is subtle and expensive. If a user told your agent in February that they worked at Company A, and in July that they moved to Company B, both statements are highly similar to a question about their employer, and the older one can win. A temporal graph tracks facts across time and records the February statement as superseded, so it stops competing with the current one. If your agent runs long enough for user facts to change — employment, preferences, plan tier, team membership — this is the difference between memory being an asset and being a source of confidently wrong answers.
Is there a free tier or trial for Zep?
No free hosted tier. The entry point is Flex at $125 per month, which is the single biggest practical objection to the product. What exists instead is Graphiti, the open-source core, at 20,000-plus GitHub stars — you can run the temporal graph engine yourself to evaluate whether the architecture fits before paying for the managed Context Lake and Context Graph Engine around it. That is a genuine evaluation path, but it is an engineering project rather than a signup. Budget a week of someone's time if you want to de-risk the $125 commitment properly, or accept the first month as the cost of the trial.
How do Zep's credits actually get consumed?
Credits are the metering unit across ingestion and retrieval, and the honest answer is that you cannot compute your monthly cost from the pricing page alone — you have to model consumption against your own traffic. What the plan structure tells you is the shape: 50,000 credits on Flex, 200,000 on Flex Plus, additional at $25 per 10,000 in both cases, which means overage is priced at the same per-credit rate as the base allowance rather than at a penalty. Combined with 30-day rollover and auto top-up at 20%, the model tolerates spiky usage better than a hard monthly reset would. The practical advice is to run a fortnight on Flex with production-shaped traffic and extrapolate, rather than trying to predict credit burn in a spreadsheet.
How does Zep compare to Mem0?
Different theories of the problem and very different price floors. Mem0 is a drop-in memory layer with a working free tier, a $19 first paid tier, and an enormous open-source community at 62,000-plus stars; it meters add and retrieval requests separately and adds graph memory at its $249 Pro tier. Zep is temporal-graph-first from the ground up, quotes sub-200ms retrieval, publishes LoCoMo and LongMemEval benchmark results, and starts at $125 with no free tier. Choose Mem0 if you are still validating that your agent needs memory at all, or if starting free matters. Choose Zep if you already know memory is core, your agent runs long enough for facts to go stale, and you need SOC 2, a HIPAA BAA or VPC deployment on a defined path.
Is Zep suitable for a regulated industry?
At the Enterprise tier, yes, and unusually clearly so for a company at this stage. SOC 2 Type II is in place, a HIPAA BAA is available, audit and API logs are retained for a year rather than a day, and deployment can be into your own VPC rather than Zep's cloud. Those are the four things a healthcare or financial-services security review will ask about, and having all four documented rather than on a roadmap is what separates a viable vendor from a promising one. The caveat is that none of it is available on Flex — one day of API log retention will not survive an audit — so a regulated buyer should treat Enterprise as the only real option and price accordingly. If you are venture-funded, ask about emerging-company pricing before assuming Enterprise is out of reach.
Related Reading
More on the infrastructure layer underneath production AI agents.
ChatGPT already recommends Zep. Does it recommend yours?
If you're building in AI agent memory tools, run a free AI-visibility scan on your own product — we ask ChatGPT across 5 prompt angles and score how often you get named. ~30 seconds, no signup, no card.
Affiliate disclosure: Some links on this page are affiliate links. If you sign up through them, AISO Tools may earn a commission at no extra cost to you. This never affects our rankings or reviews.
📬 Get the best new AI tools delivered weekly
One concise email with fresh launches, trending picks, and featured standouts.
Join thousands of professionals who discover the best AI tools every week. No spam — unsubscribe anytime.