Recall and Zep compared
Last updated: 13 October 2026
Zep builds a temporal knowledge graph for each user from chat messages and business data, and returns a block of relevant context for your agent's prompt. It is a hosted service; its graph engine, Graphiti, is open source.
Every fact about Zep on this page is from Zep's own public pages, linked under Sources, as we read them on 13 October 2026. What we could not check there is left out. Tell us at support@recallmem.dev if something has changed.
At a glance
| Recall | Zep | |
|---|---|---|
| What it is | A memory API. It learns what is true about each person from your agent's conversations and answers questions about them, with sources, in one model call. | A context platform that keeps a temporal knowledge graph per user (and shared graphs), built from messages and business data, and assembles a context block from it.12 |
| Hosted service | Yes. Data is stored in Frankfurt (EU). | Yes, Zep Cloud. Enterprise adds your own encryption keys (BYOK) or Zep running in your own cloud (BYOC).3 |
| Self-hosting and licence | Yes. The server is source-available under the FSL-1.1-ALv2 (each version becomes Apache 2.0 after two years); the SDKs and integrations are MIT. Self-hosting, source. | Zep Community Edition is no longer maintained (since April 2025; the repository stays Apache 2.0). Graphiti, the graph framework Zep is built on, is Apache 2.0.4 |
| Pricing | Pay as you go from prepaid credit: learning $1.75 per 1M message tokens, questions $0.0008 to $0.25 by level. Context, search, storage and the night shift are included. Pricing. | Credits per episode sent: free, Flex $125 a month, Flex Plus $375 a month, Enterprise custom. Retrieval and storage use no credits.3 |
| Free to start | $5 of credit once, no card needed. | 10,000 credits a month, with lower limits and lower processing priority.3 |
| API | The v3 memory REST API, which Honcho's clients also speak; TypeScript and Python SDKs, an MCP server, and plugins for Claude Code and Codex. | Zep's own API: users, threads, messages and graphs. Python, TypeScript and Go SDKs; a Memory MCP server, with seats by plan.123 Not compatible with Recall's API. |
Pricing
Zep charges credits for what you send: an episode (a chat message, a JSON payload or a block of text) uses 1 credit per 350 bytes or part of it. Retrieval, storage, threads, users and graph storage use none.3 Recall charges per message token learned and per question; context, search and storage are included.
| Zep plan3 | Price a month | Credits included | More credits |
|---|---|---|---|
| Free | $0 | 10,000 | No top-up |
| Flex | $125 | 50,000 | $25 per 10,000 |
| Flex Plus (adds webhooks, analytics, custom extraction instructions) | $375 | 200,000 | $75 per 40,000 |
| Enterprise (lists SOC 2 Type II, a HIPAA BAA, BYOK and BYOC) | Custom | Custom | Negotiated |
Recall: learning $1.75 per 1M message tokens (a message counts as at least 50 tokens), questions from $0.0008 to $0.25 by level, no monthly fee, $5 of credit to start. Zep returns context for your own model to answer from; the closest Recall calls, context and search, are included in its price.
Published benchmark results
Each figure is as its maker published it, with its own answer model, judge and harness. None is a rerun, and a different judge alone can move a score by several points, so read the last column before comparing. Recall's runs and method are on Evals.
| Benchmark | Recall | Zep | How each was graded | Compare directly? |
|---|---|---|---|---|
| LongMemEval | 91.0% (455 of 500) on LongMemEval-S | 90.2% (451 of 500)5 | Recall: LongMemEval's official judge, gpt-4o-2024-08-06 with the benchmark's own prompts. Zep: answers by gpt-5.4 (reasoning medium), judged by gpt-5.4 with chain-of-thought grading; the page does not say which LongMemEval variant.5 | No: a different judge, and the variant is not named. |
| LoCoMo | Not run in full. On a 75-question sample: 60.0% with a strict judge, 79.2% with a lenient one | 94.7% (1,459 of 1,540)5 | Zep: the same models as its LongMemEval run.5 | No: Recall has no full LoCoMo run. |
Zep's research page also reports retrieval latency and context size per question.5 We found no BEAM result published by Zep.
Hosting and licences
Hosted. Zep Cloud, with options on Enterprise to bring your own encryption keys or to run Zep inside your own cloud account. Zep states it is SOC 2 Type II certified and offers HIPAA BAAs on Enterprise.3 Recall runs a hosted service with data stored in Frankfurt (EU).
Self-hosted. Zep stopped maintaining Zep Community Edition in April 2025. Graphiti, Zep's temporal knowledge graph framework, is open source under Apache 2.0.4 Recall's whole server can be self-hosted, source-available under the Functional Source License (each version Apache 2.0 after two years).
API
Zep has its own API: each user has a graph and a thread history; messages added to a thread are ingested into the user's graph; facts are edges, and a fact that a newer one invalidates keeps the time it became invalid.1 It is not compatible with Recall's API. A Zep user maps to a Recall peer and a thread to a session. To move, send the conversations you kept to Recall's messages endpoint and Recall learns from them.
When to choose Zep
- You need a knowledge graph of entities and relationships, with your own entity and edge types, built from business data as well as chat.
- You need compliance paperwork or deployment options that Recall does not offer: SOC 2 Type II, a HIPAA BAA, your own encryption keys, or running in your own cloud (Enterprise).
- You want to pay per message sent, with retrieval free.
- You want a Go SDK.
When Recall fits better
- You want Recall to write the answer, with citations, a confidence and an explicit "I don't know", in one model call.
- You want to self-host the whole memory service, not only a graph framework.
- You want pay as you go without a monthly plan, data stored in the EU, or the v3 memory API that Honcho's clients also speak.
See all comparisons, or try Recall with $5 of free credit: quickstart.
Sources
Read on 13 October 2026.