Zep

Value Proposition & Features

• Zep’s Agent Memory service extracts entities and facts from conversations and business data, stores them in a temporal knowledge graph, and serves relevant context back to agents at request time. ^{[k9jjmf] [r1ze78] [7cee07]}
• The Context Graph / Context Lake provides bi-temporal edges with validity intervals and fact invalidation, enabling queries like “what was true at time X” versus “what is true now,” and auditing how knowledge evolved over time. ^{[it7zro] [s34gjv] [7cee07]} [^ear82x]
• Graphiti, the open‑source engine behind Zep, powers hybrid retrieval (embeddings + BM25 + graph traversal) with sub‑200ms reads, and can be used standalone with a bring‑your‑own graph database. ^{[k9jjmf] [r1ze78] [s34gjv] [4kbl3v]}

• Temporal Context Graph (bi‑temporal knowledge graph) — facts stored as edges with event time, ingestion time, and validity intervals, supporting “what used to be true” queries and audit trails. ^{[it7zro] [s34gjv] [7cee07]}
• Hybrid Retrieval (search + graph traversal) — recall queries run semantic similarity and BM25 search over nodes and edges, fuse results, then traverse subgraphs via BFS and related algorithms to assemble context efficiently. ^{[k9jjmf] [s34gjv]}
• Observations & Pattern Detection — Zep builds cross‑conversation patterns as “Observations,” durable, evidence‑backed summaries of what the graph shows about users or problems rather than just episodic facts. ^{[85fffn] [tdli9z] [19qz3y]}
• Provenance & Lineage Tracking — every derived fact links back to raw episodes, enabling veracity checks, source attribution for LLM‑synthesized answers, and compliance use cases such as GDPR deletion. ^[19qz3y] [^ear82x] [^xbk61n]
• Managed Cloud Context Graph Engine — Zep Cloud includes a proprietary graph runtime with SLAs, sub‑200ms retrieval, and no need for a third‑party graph database, unlike Graphiti which requires BYO graph store. ^{[r1ze78] [4kbl3v]}
• Agent Framework Integrations & MCP — integrations and guides for Claude Code, Codex, Cursor, Strands, Eve, and Memory MCP let agents use Zep for long‑term memory with minimal setup. ^{[5cja6r] [mscs7u]} [^eg3e7p] [^dnww73]
• Enterprise Governance & Access Control — attribute‑based access control on graph edges and episodes, user/group policies, and controls for write permissions on MCP connections. [^u5eosb] [^f1avlu] [^dnww73]
• Strict Ontology & Ingestion Pipeline Support — strict ontology flags and ingestion guides ensure only configured entity/edge types are added, with batch APIs and data‑prep guides for production pipelines. [^2disjl] [^w73flu]

Product Roadmap / Announcements

• 2026-08-13 — Documentation updates: Account Owners can configure enterprise SSO without sharing IdP credentials, custom OpenAI base URLs (e.g., regional hosts), new Strands and Eve integration guides, and documentation for the zep-strands package for Strands. [^dnww73]
• 2026-08-04 — Memory MCP connections now allow writes by default; administrators can switch to read-only, and connections provisioning users just‑in‑time now recover automatically after user deletion, with improved name display and documentation of how graph search interprets relative calendar phrases. [^f1avlu]
• 2026-08-03 — A “Memory for Agent Frameworks” landing page was added, the Ecosystem section renamed to Agent Frameworks, and the “Evaluate Zep for Your Use Case” guide rewritten to cover evaluation workflows including ingestion, search scopes, and graph inspection. [^6sul99]
• 2026-07-30 — Batch API jobs now send a single completion webhook to avoid duplicate notifications; strict ontology documented for limiting extraction to configured types, and ingestion pipeline guides added. [^2disjl] [^w73flu]
• 2026-07-29 — Graph node and edge lists can be ordered by valid time (valid_at), and company email signups receive a credit‑limited 30‑day Enterprise trial before reverting to Free; ABAC constraints are enforced on updates/deletes of graph artifacts. [^u5eosb] ^[42qbfq] [^4ecgew]
• 2026-07-18 — New “Implement Zep with agents” page under Developer Tools covering the Build with Zep plugin for Claude Code and Codex, plus onboarding and UI tweaks in the graph dialog. [^eg3e7p] ^[bvkq40]
• 2026-07-16 — Blog post detailing benchmarking of NVIDIA Nemotron 3 Embed 1B for agent memory, showing it as Zep’s best‑performing embedding on their production recall workload. ^[k9jjmf] [^fphk7u] [^r5rfm9]
• 2026-07-30 — “Build with Zep plugin: Claude Code, Codex, Cursor” announcement: one plugin bundle and MCP documentation server (zep-docs) that lets coding agents design Zep implementations via guided steps. ^{[5cja6r] [o75lsm] [mscs7u]}

Recent Developments (last 90 days)

• Self-host model change (2026-08-07) — Independent analysis reports that Zep’s Community Edition is deprecated and getzep/zep now hosts examples and integrations, making Zep a hosted‑only product; self‑hosting is now limited to building on Graphiti rather than deploying Zep Cloud itself. [^9v5f60] [^97r09g]
• Graph traversal & FalkorDB local mode (2026-07-27–08-02) — Graphiti releases add FalkorDB bug fixes and support for fully local operation (embedded FalkorDB, Ollama, no server/cloud LLM), while Zep Cloud exposes context graph traversal APIs for walking neighbors and multi‑hop subgraphs. [^c9u8ih] [^phtu56] ^[d3o7bu] [^yv02vx] ^[r1ze78]
• Embedding model upgrade (2026-07-16) — Zep benchmarks multiple embedding models and concludes Nemotron 3 Embed 1B is the strongest tested on its agent‑memory workload, with statistically significant improvements in Recall@10 and a quantized serving path. ^[k9jjmf] [^fphk7u] [^r5rfm9]
• Plugin & framework integrations (late July–early August) — Zep ships a unified plugin for Claude Code, Codex, and Cursor and documents Strands and Eve integration guides to add long‑term memory to these agent frameworks. ^{[5cja6r] [mscs7u]} [^dnww73]
• Pricing tier for emerging companies (2026-07-21–08-02) — Zep introduces “Enterprise memory priced for emerging companies” at $13,000 for the first year, targeting companies that have raised $1M–$10M, with more than $40,000 in discounts over a three‑year term. [^9csuah] [^6djpgd] [^zzw4r7] [^g6uj3a]

History and Origin Story

Fundraising History

• Y Combinator [^83i6as]

Notable Team Members

Market Sizing

Category, Market Size, and Category Growth

Pricing

Revenue Trajectory Estimates

Competitive Landscape

Who it’s for, who it’s not for

Viable Alternatives

• Mem0 — Vector‑centric agent memory with both managed cloud and open‑source self‑host options, suitable for teams prioritizing simple, fast memory integration and hard self‑host/data‑residency requirements. [^9v5f60] ^{[hu5yt4] [37gaf1] [kv8ehv]} [^bs8o36] [^1uyakk]
• Letta (MemGPT) — Agent runtime with hierarchical memory management, open source and self‑hostable, attractive for teams wanting an agent framework with built‑in tiered memory rather than a separate memory service. ^{[7cee07] [jr7g1a] [kv8ehv] [37gaf1]}
• Supermemory — Vector‑graph memory platform emphasizing a simple universal memory API and free local self‑hosting, good for rapid deployment and mixed semantic/keyword retrieval without complex temporal modeling. ^{[7cee07] [37gaf1] [9ppgt9]}
• Cognee — Knowledge-graph plus embeddings hybrid retrieval, with managed or self‑host deployments, targeting document/data-heavy pipelines rather than conversational memory alone. ^{[g2ogtg] [7cee07] [qfsbk9]}
• HydraDB — Graph‑native context layer database positioned as an alternative to Mem0 and Zep for production AI agents requiring persistent context and complex state tracking. ^{[kv8ehv] [zy7zu8]} [^sm5hqw]

Competitor Table

Sources