Zep

Name: Zep
Brand: Zep

AI Native

Zep is an agent-memory infrastructure layer that provides durable, cross-session temporal context graphs by ingesting chat history, user interactions, and structured business data to return provenance-aware context and reduce token usage per agent turn.

Est 2023|Valuation $6.8M|Raised $2.3M|27 Ref Customers

LinkedIn 3K|Web traffic 45K|FTEs 8|HQ San Francisco, US

MarketsAgent Memory Infrastructure

Published Oct 15, 2026

01Common complaints

Company sizes or stages where this product is overkill or insufficient
- Early prototypes and single-agent demos:
  - A simple database plus prompt strategy and short-term memory may satisfy requirements without adopting a full agent-memory platform.
Industries or use cases where competitors are clearly superior
- Workflows requiring fully inspectable, version-controlled memory artifacts:
  - File-based or document-centered memory approaches may be preferred when inspectability and version control are primary requirements.
Technical requirements that this product cannot meet
- If an organization must keep all data on-prem with no cloud service dependencies:
  - Zep’s managed cloud features and analytics may not be fully available unless Bring Your Own Cloud (BYOC) or self-hosting options meet the organization’s constraints.
  - Self-hosting Graphiti alone may omit the vendor’s managed governance and analytics capabilities.
Budget constraints that make alternatives more appropriate
- High-volume ingestion workloads (event-heavy telemetry, tool traces, verbose logs) can drive credit usage under the vendor’s pricing model.
- DIY approaches can reduce vendor spend but increase engineering costs.
Specific workflows or integrations where the product falls short
- If an agent does not reason about time, the platform’s temporal features may provide limited value while still adding operational complexity.
Red flags in the buying process
- If guaranteed rate limits and strict SLAs are required, Enterprise-level negotiation may be necessary, which can lengthen procurement timelines.
Current limitations acknowledged by the vendor
- Vendor materials do not enumerate a comprehensive list of limitations; public messaging primarily emphasizes capabilities.
Features on the roadmap (not yet available)
- No formal roadmap commitments with dates were available in the accessible sources for this report.
Scale limitations
- The vendor reports retrieval latency remaining under 200ms regardless of graph size; independent public verification of this claim is limited and should be validated in a proof-of-concept.
Security or compliance gaps
- ISO certifications (e.g., ISO 27001/27017/27018) were not confirmed in accessible public sources for this report.
- PCI data is explicitly listed as prohibited in the vendor Terms unless those Terms change.
Technical debt or architectural limitations mentioned in reviews
- Community themes include the complexity and maintenance overhead of graph-based memory and incorrect memories if extraction/selection is not tuned and evaluated.
API limitations or integration constraints
- Framework integrations exist, but teams must design ingestion policies and evaluation harnesses.
- Self-hosting Graphiti may require building operational capabilities (dashboards, analytics) that Zep Cloud includes as part of the managed offering.

- Company sizes or stages where this product is overkill or insufficient - Early prototypes and single-agent demos: - A simple database plus prompt strategy and short-term memory may satisfy requirements without adopting a full agent-memory platform. - Industries or use cases where competitors are clearly superior - Workflows requiring fully inspectable, version-controlled memory artifacts: - File-based or document-centered memory approaches may be preferred when inspectability and version control are primary requirements. - Technical requirements that this product cannot meet - If an organization must keep all data on-prem with no cloud service dependencies: - Zep’s managed cloud features and analytics may not be fully available unless Bring Your Own Cloud (BYOC) or self-hosting options meet the organization’s constraints. - Self-hosting Graphiti alone may omit the vendor’s managed governance and analytics capabilities. - Budget constraints that make alternatives more appropriate - High-volume ingestion workloads (event-heavy telemetry, tool traces, verbose logs) can drive credit usage under the vendor’s pricing model. - DIY approaches can reduce vendor spend but increase engineering costs. - Specific workflows or integrations where the product falls short - If an agent does not reason about time, the platform’s temporal features may provide limited value while still adding operational complexity. - Red flags in the buying process - If guaranteed rate limits and strict SLAs are required, Enterprise-level negotiation may be necessary, which can lengthen procurement timelines. - Current limitations acknowledged by the vendor - Vendor materials do not enumerate a comprehensive list of limitations; public messaging primarily emphasizes capabilities. - Features on the roadmap (not yet available) - No formal roadmap commitments with dates were available in the accessible sources for this report. - Scale limitations - The vendor reports retrieval latency remaining under 200ms regardless of graph size; independent public verification of this claim is limited and should be validated in a proof-of-concept. - Security or compliance gaps - ISO certifications (e.g., ISO 27001/27017/27018) were not confirmed in accessible public sources for this report. - PCI data is explicitly listed as prohibited in the vendor Terms unless those Terms change. - Technical debt or architectural limitations mentioned in reviews - Community themes include the complexity and maintenance overhead of graph-based memory and incorrect memories if extraction/selection is not tuned and evaluated. - API limitations or integration constraints - Framework integrations exist, but teams must design ingestion policies and evaluation harnesses. - Self-hosting Graphiti may require building operational capabilities (dashboards, analytics) that Zep Cloud includes as part of the managed offering.

Deal disqualifiers: what makes a POOR fit

The buyer only requires a simple user profile store and a small set of canonical facts.
The organization cannot tolerate additional platform complexity.
The procurement process requires abundant third-party enterprise reviews and artifacts that are currently limited or unclear in public sources.

Why some customers left for competitors (churn reasons)

Direct vendor-specific churn stories were not abundant in the accessible sources.
Indirect reasons cited in community discussions include:
- Overhead and complexity relative to simpler RAG or database-based approaches.
- Memory systems remembering incorrect or irrelevant items due to automated extraction/selection decisions.
- Cost concerns when memory ingestion volume grows.

02Why they buy

Competitive advantages observed in community feedback and vendor materials
- Community discussions recognize temporal graphs as useful when time-sensitive correctness matters.
- The platform’s design emphasizes provenance tracing and retention controls useful for regulated deployments.
Unique features and capabilities competitors commonly lack
- Temporal validity mechanisms to model changes over time and invalidate outdated facts while preserving history
- Provenance tracing to connect facts back to source episodes
- Enterprise governance features (ABAC, retention, audit) integrated into the platform per vendor documentation
Differentiation from competitors based on reviews
- Public review footprint for the specific agent-memory product is limited; most differentiation signals are architecture-driven rather than review-driven.
Areas where competitors clearly have advantages
- Simpler, lower-overhead adoption paths (often cited for Mem0 or DIY approaches)
- Agent runtime-centric memory patterns (often cited for Letta/MemGPT)

- Competitive advantages observed in community feedback and vendor materials - Community discussions recognize temporal graphs as useful when time-sensitive correctness matters. - The platform’s design emphasizes provenance tracing and retention controls useful for regulated deployments. - Unique features and capabilities competitors commonly lack - Temporal validity mechanisms to model changes over time and invalidate outdated facts while preserving history - Provenance tracing to connect facts back to source episodes - Enterprise governance features (ABAC, retention, audit) integrated into the platform per vendor documentation - Differentiation from competitors based on reviews - Public review footprint for the specific agent-memory product is limited; most differentiation signals are architecture-driven rather than review-driven. - Areas where competitors clearly have advantages - Simpler, lower-overhead adoption paths (often cited for Mem0 or DIY approaches) - Agent runtime-centric memory patterns (often cited for Letta/MemGPT)

Why customers chose this product over alternatives (inferred from discussions)

Temporal validity and contradiction handling (handling “what’s true now” vs “what was true before”) is often cited as a differentiator relative to vector-only memory approaches.
Governance posture (ABAC, retention, audit, provenance) is a decision factor for enterprise and regulated use cases.
Integration and convenience: the vendor provides memory APIs, SDKs, and framework integrations (LangGraph, AG2, Google ADK, Microsoft Agent Framework), plus the Graphiti server.

- Temporal validity and contradiction handling (handling “what’s true now” vs “what was true before”) is often cited as a differentiator relative to vector-only memory approaches. - Governance posture (ABAC, retention, audit, provenance) is a decision factor for enterprise and regulated use cases. - Integration and convenience: the vendor provides memory APIs, SDKs, and framework integrations (LangGraph, AG2, Google ADK, Microsoft Agent Framework), plus the Graphiti server.

03Field reports

Important data-quality note (brand confusion)

Public review platforms frequently conflate “Zep” (the agent-memory product marketed on the vendor's website) with other products named “ZEP” (notably time tracking/resource/invoicing products and unrelated consumer items). Review pages that clearly describe time tracking/ERP functionality should be treated as unrelated to the agent-memory product.
As a result, product-specific review coverage for the agent-memory offering on major review platforms appears limited.

- Public review platforms frequently conflate “Zep” (the agent-memory product marketed on the vendor's website) with other products named “ZEP” (notably time tracking/resource/invoicing products and unrelated consumer items). Review pages that clearly describe time tracking/ERP functionality should be treated as unrelated to the agent-memory product. - As a result, product-specific review coverage for the agent-memory offering on major review platforms appears limited.

Review platforms

G2
- A G2 listing exists for “ZEP” with a very small review count. The product mapping on review platforms may not reliably distinguish the agent-memory product from other products with the same name.
TrustRadius
- TrustRadius shows a listing for “ZEP” with no reviews on at least one page and a description oriented to time tracking/business management, suggesting potential brand confusion.
Capterra
- Capterra’s “ZEP” listing includes many reviews but describes resource planning/time tracking functionality unrelated to the agent-memory product under review.

- G2 - A G2 listing exists for “ZEP” with a very small review count. The product mapping on review platforms may not reliably distinguish the agent-memory product from other products with the same name. - TrustRadius - TrustRadius shows a listing for “ZEP” with no reviews on at least one page and a description oriented to time tracking/business management, suggesting potential brand confusion. - Capterra - Capterra’s “ZEP” listing includes many reviews but describes resource planning/time tracking functionality unrelated to the agent-memory product under review.

Community/social (more directly relevant to the vendor's agent-memory offering)

Reddit and Hacker News discussions are more directly focused on agent memory and frequently mention Zep alongside Mem0, Letta, Graphiti, and DIY approaches.
Representative paraphrased comments from community threads (technical and anecdotal):
- Positive/neutral-technical:
  - One thread noted that Zep uses a temporal knowledge graph and can represent not only what happened but when it happened and how facts change over time.
  - An informal post expressed strong approval for Graphiti in colloquial terms.
- Critical / skeptical:
  - Community members observed that Zep’s temporal graph can be useful, but only if an agent’s architecture reasons about time; this introduces adoption complexity.
  - Some users argued that small-scale DIY approaches (for example, a SQLite-based solution with full-text search) can be adequate for personal-scale use cases and simpler to maintain.
  - Discussions on Hacker News raised that graphs handle structure well but add complexity and maintenance overhead compared with simpler memory approaches.

- Reddit and Hacker News discussions are more directly focused on agent memory and frequently mention Zep alongside Mem0, Letta, Graphiti, and DIY approaches. - Representative paraphrased comments from community threads (technical and anecdotal): - Positive/neutral-technical: - One thread noted that Zep uses a temporal knowledge graph and can represent not only what happened but when it happened and how facts change over time. - An informal post expressed strong approval for Graphiti in colloquial terms. - Critical / skeptical: - Community members observed that Zep’s temporal graph can be useful, but only if an agent’s architecture reasons about time; this introduces adoption complexity. - Some users argued that small-scale DIY approaches (for example, a SQLite-based solution with full-text search) can be adequate for personal-scale use cases and simpler to maintain. - Discussions on Hacker News raised that graphs handle structure well but add complexity and maintenance overhead compared with simpler memory approaches.

Specific outcomes customers achieved (with metrics when available)

The vendor publishes benchmark metrics (LoCoMo, LongMemEval) reporting accuracy and retrieval latency figures.
Independent validation is mixed: some independent research reports materially lower Zep accuracy on LoCoMo under a matched evaluation setup, indicating that benchmark outcomes are sensitive to methodology and model selection.

Deal-breakers and reasons customers didn’t purchase

Potential deal-breakers observed in community patterns include:
- Preference for fully self-hosted, minimal managed-service dependencies.
- Teams that primarily need stable identity context (names, addresses, preferences) often prefer structured databases and explicit rules over memory extracted from chat.
- Organizations with strict data residency requirements may require BYOC; if that deployment option is not available at an affordable tier, they may not proceed.

- Potential deal-breakers observed in community patterns include: - Preference for fully self-hosted, minimal managed-service dependencies. - Teams that primarily need stable identity context (names, addresses, preferences) often prefer structured databases and explicit rules over memory extracted from chat. - Organizations with strict data residency requirements may require BYOC; if that deployment option is not available at an affordable tier, they may not proceed.

Case Studies

Vendor-published and vendor-curated evidence
- The vendor highlights customer names and includes role-based quotes (e.g., from CTOs or VPs). These are vendor-provided testimonials and are not independently verified as production deployments in this report.
- The vendor references an industry market intelligence report describing Zep’s approach and potential strategic role.
Critical evaluation
- Benchmark claims
  - The vendor publishes benchmark results on LoCoMo and LongMemEval with reported accuracy and latency metrics.
  - Third-party research indicates that benchmark outcomes are highly sensitive to evaluation design; independent research reports substantially different results when matching models and prompts, suggesting vendor-published numbers should be treated as directional.
  - Recommendation: request full methodology, model versions, prompts, and reproducibility artifacts when evaluating benchmark claims.
What’s missing for a robust case study
- Independent, end-to-end production case studies with:
  - Baseline vs post-adoption outcomes
  - Documented failure modes encountered (wrong recall, privacy incidents, latency spikes)
  - Measured token reduction, cost per successful task, and user satisfaction
  - Audit outcomes for regulated deployments

- Vendor-published and vendor-curated evidence - The vendor highlights customer names and includes role-based quotes (e.g., from CTOs or VPs). These are vendor-provided testimonials and are not independently verified as production deployments in this report. - The vendor references an industry market intelligence report describing Zep’s approach and potential strategic role. - Critical evaluation - Benchmark claims - The vendor publishes benchmark results on LoCoMo and LongMemEval with reported accuracy and latency metrics. - Third-party research indicates that benchmark outcomes are highly sensitive to evaluation design; independent research reports substantially different results when matching models and prompts, suggesting vendor-published numbers should be treated as directional. - Recommendation: request full methodology, model versions, prompts, and reproducibility artifacts when evaluating benchmark claims. - What’s missing for a robust case study - Independent, end-to-end production case studies with: - Baseline vs post-adoption outcomes - Documented failure modes encountered (wrong recall, privacy incidents, latency spikes) - Measured token reduction, cost per successful task, and user satisfaction - Audit outcomes for regulated deployments

04Best fit

Company sizes

Likely best fit:
- Mid-market to enterprise teams (roughly mid-sized to large organizations) running production agents across multiple workflows.
- Smaller, well-funded AI product companies where memory quality is central to the product may also be suitable.

Industries and verticals

Strongest alignment:
- SaaS companies delivering personalized agent experiences
- Security- and compliance-driven industries (healthcare, finance) where provenance and retention controls are important

Roles and personas

Decision makers
- Head of AI/ML Platform, CTO, VP Engineering, Security/Compliance leadership
End users
- AI engineers building agents, platform engineers, applied ML engineers

Pain points and challenges the product addresses

Multi-session agents that require:
- Durable memory across time
- Correct handling of changing facts
- Governed access across many agents and users

Qualification criteria: what makes a GOOD fit

The buyer operates production agents with measurable failure modes due to missing or incorrect long-term context.
The organization requires auditability, retention controls, and least-privilege access.
Sufficient engineering capacity exists to operationalize memory evaluation and schema design.

Real customer examples

The vendor lists customer names/logos under a “Trusted By AI Teams” heading (e.g., Samsung, Zscaler) on marketing pages, but these entries were not independently verified as production deployments in the sources accessible for this report.

05The pitch

What the product does (core value proposition in objective terms)

Zep is an agent-memory infrastructure layer designed to provide AI agents with durable, cross-session memory.
The vendor positions Zep as a temporal context-graph system that ingests multiple sources (chat history, user interactions, and structured business data) and returns context intended to reduce token usage and provide relevant information for each agent turn.
Zep’s product messaging focuses on:
- Temporal facts (what was true when, what is true now)
- Graph-based memory (referred to by the vendor as “Context Graphs” and “Context Lake”)
- Governance features (access control, retention, audit, provenance)
- Deployment options intended for enterprise trust boundaries (cloud, BYOK, BYOC)

- Zep is an agent-memory infrastructure layer designed to provide AI agents with durable, cross-session memory. - The vendor positions Zep as a temporal context-graph system that ingests multiple sources (chat history, user interactions, and structured business data) and returns context intended to reduce token usage and provide relevant information for each agent turn. - Zep’s product messaging focuses on: - Temporal facts (what was true when, what is true now) - Graph-based memory (referred to by the vendor as “Context Graphs” and “Context Lake”) - Governance features (access control, retention, audit, provenance) - Deployment options intended for enterprise trust boundaries (cloud, BYOK, BYOC)

Target market and market positioning

Primary target is B2B customers: enterprise teams building production AI agents that require governed, long-term memory across many agents and many users.
Secondary audience: developers building agent-based applications who want rapid integration via SDKs.
The vendor markets to regulated environments via claims of SOC 2 Type II certification and an available HIPAA BAA for Enterprise customers, and emphasizes governance integrated into the platform.

Get a comprehensive analysis of Zep including market position, competitive landscape, adoption trends, and peer benchmarks.

Download full report →