2026年5月更新
Zep is a long-term memory backend for AI agents. Its core differentiator is the temporal knowledge graph: as conversations flow in, Zep automatically extracts entities (people, projects, products, concepts), tracks their attributes over time, and lets the agent query "what did we know about X on date Y" or "what changed about X between then and now". The Graphiti graph engine powers this; the SDK and managed cloud product wrap it.
Zep is positioned for engineering teams building AI agents at scale — typically B2B SaaS products where the AI handles complex multi-turn interactions and the team values typed graph reasoning over flat vector retrieval. Pricing is mostly enterprise / custom; an open-source community edition exists for self-hosting.
Alma is a complete persistent memory product. End users sign up at alma.olivares.ai, chat with Anthropic Claude (Haiku, Sonnet, Opus 4.7 with 1M-token context) and the memory layer captures facts, preferences and decisions automatically. The Soul Engine handles AI identity. Image, Video and Music studios are bundled in Pro and Max. Developers can also use the MCP server, the JavaScript SDK and the REST API on any paid plan.
Memory is structured into three layers (memories, episodes, procedures) with five-factor scoring. Paid plans start at $14/mo (Starter) with full Soul Engine access.
Zep's temporal knowledge graph is its strongest feature. Entities are first-class citizens; you can ask "what is project X?", "who works on project X?", "when did the deadline change?" and Zep traverses the graph to answer. This is genuinely powerful when the AI needs to reason about relationships across many entities and how they evolve over time — typical of customer-support AI, research agents and B2B workflow automation.
Alma's 3-layer model is opinionated about shape: memories (facts and preferences), episodes (conversation summaries) and procedures (learned workflows). Entity extraction runs heuristically on every memory create (no extra LLM cost on read) and feeds an internal knowledge graph, but the user-facing model is the three layers, not the graph itself. Retrieval is hybrid (semantic + keyword) with five-factor scoring. The trade-off: less explicit graph reasoning, but a much simpler mental model and a working assembled prompt in under 100ms with no graph query language to learn.
Zep is infrastructure: an engineering team integrates it into their AI product. Most users never see Zep's name; they see your product, which happens to remember them. Alma is a product: end users sign up directly, the Alma brand is the experience, and developers are an additional audience served by the SDK and MCP server.
If your team is building an AI agent with sophisticated multi-entity reasoning at the heart of it, Zep was designed for exactly that. If you (or your team) want to use AI with memory across the tools you already work in — Cursor, Claude Desktop, VSCode — Alma covers that without writing any integration code.
| Feature | Zep | Alma |
|---|---|---|
| Target user | Developers / engineering teams building AI agents | End users + developers |
| Memory model | Temporal knowledge graph with entity extraction (Graphiti) | 3-layer (memories + episodes + procedures) with internal entity graph |
| Web app / chat | No — API and admin dashboard only | Yes — full chat with streaming, tools, file attachments |
| AI identity | Not provided — bring your own | Soul Engine (13 versioned identity blocks) |
| Temporal awareness | First-class — track entity changes across time | Episodes + memory timestamps + importance decay (no graph query language) |
| Context assembly | Bring your own — graph queries return data, you build the prompt | Built-in — assembled system prompt <100ms, 5-factor scoring |
| Integrations | SDK, REST API, framework adapters | MCP (Claude Desktop / Cursor / Windsurf), VSCode, JS SDK, REST API |
| Creative tools | None | Image / Video / Music studios in Pro and Max |
| BYOK | N/A | Every plan, incl. Free — Anthropic, Replicate, Leonardo, ElevenLabs |
| Setup complexity | High — graph DB concepts, integration work | Sign up and chat — zero infrastructure for users |
| Pricing | Open-source community edition · cloud / enterprise (custom) | Starter $14/mo · Pro $29/mo · Max $99/mo |
Zep's open-source community edition runs free if you self-host the graph database and write the integration code. The hosted cloud and enterprise tiers are quoted on demand and aimed at teams shipping AI agents at scale. Alma's pricing is straightforward and end-user-friendly: Starter $14/mo (unlimited memories, Haiku + Sonnet), Pro $29/mo (adds Opus 4.7 with 1M-token context, unlimited environments and creative studios), Max $99/mo (largest monthly budget — API, MCP server, VSCode extension and SDK come with every paid plan; BYOK works on every plan, including Free). Annual billing is "pay 10 months, get 12" on every paid plan.
End-user memory across tools. A user wants AI with memory across Claude Desktop, Cursor, VSCode and the web. Zep has no end-user product — it is infrastructure to embed in someone else's product. Alma is the end-user product itself: sign up, connect the MCP server, and every supported client has memory in five minutes.
B2B customer-support agent with temporal reasoning. An engineering team is building an AI agent that handles support tickets and needs to reason about "what did we know about account X on date Y" or "when did the contract terms change". Zep's temporal knowledge graph (Graphiti) is purpose-built for this — entities are first-class citizens, attributes are tracked across time, graph queries answer the temporal questions naturally. Alma's typed 3-layer model can answer "what do we know about X" via search, but does not expose explicit temporal graph queries — that is Zep's differentiated strength.
Solo developer wanting memory in their own AI side project. A developer is building a small AI tool and wants persistent memory without standing up a graph database, running migrations, and writing graph query code. Alma's SDK (on any paid plan, from $14/mo) provides a complete cognitive layer (memory + Soul + context assembly) with an HTTP API. Zep's self-hosted community edition is also free for this case but expects engineering investment before the first feature ships.
Choose Zep if you are an engineering team building an AI agent product where temporal entity reasoning is core to the experience — a customer-support AI tracking accounts, a research agent following hypotheses across weeks, an enterprise workflow automation that needs to know exactly when a fact changed and why. The graph model pays off when "show me the state of entity X on date Y" is a real query you need to answer. You will write integration code and probably a wrapper UI; that is the deal.
Choose Alma if you want to use AI memory rather than build infrastructure. Sign up, start chatting, and memory works. If you also work in Cursor / Claude Desktop / VSCode / Windsurf, install the MCP server (free) and the same memory follows you. Choose Alma if you want a structured AI identity (Soul Engine), bundled creative studios or an affordable single-user subscription. Developers who want to embed Alma into their own apps get the SDK, MCP server and REST API on any paid plan (from $14/mo) — a complete cognitive layer instead of just a memory backend.
Does Alma have temporal awareness? Yes, but with a different model. Memories carry creation and last-access timestamps; episodes are time-ordered conversation summaries; importance decay (5-factor scoring includes recency 10%) lets older memories fade naturally. There is no graph query language for "state at time T" — that is Zep's differentiated feature, not Alma's.
Does Alma have an entity knowledge graph? Yes, internally. Entity extraction runs heuristically on every memory create (no extra LLM cost on read) and feeds the graph used at retrieval time. It is not user-queryable as a graph today; the user-facing model stays "memories / episodes / procedures".
Can Alma replace Zep in my AI agent product? Sometimes. If your agent is end-user-facing and the memory needs are "remember preferences, decisions and context across sessions", Alma's SDK + REST API (on any paid plan) give you what you need without building a wrapper. If you need explicit temporal graph queries, Zep is the right tool.
Is Alma open source? No. Alma is a hosted SaaS. Zep ships an open-source community edition you can self-host.
Zep is purpose-built for engineering teams shipping AI agents that need explicit temporal graph reasoning. Alma is the right tool when you want a complete memory product (chat, Soul Engine, MCP, creative studios) without infrastructure work, or when you need a cognitive layer for your own app without writing the typed memory model and Soul Engine yourself. The Starter plan ($14/mo) shows you the product end-to-end in five minutes.