Recall and Honcho compared
Last updated: 13 October 2026
Honcho, by Plastic Labs, is a memory API for agents. Recall implements Honcho's v3 REST API, so the two take the same requests and code moves between them by changing the base URL and the key.
Every fact about Honcho on this page is from Plastic Labs' 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 | Honcho | |
|---|---|---|
| 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 memory and reasoning API for agents, by Plastic Labs. Its ingestion price includes reasoning by Neuromancer, Plastic Labs' own model.1 |
| Hosted service | Yes. Data is stored in Frankfurt (EU). | Yes, at app.honcho.dev.1 |
| 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. | Yes. Open source under the AGPL-3.0. Needs Postgres with pgvector and at least one LLM provider key; the Docker setup also runs Redis for caching.23 |
| 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. | Pay as you go: ingestion $2.00 per 1M tokens, questions $0.001 to $0.50 by level. context() and dreaming are included.1 |
| Free to start | $5 of credit once, no card needed. | $5 of credit for a new account, and $5 more every month once a card is added. Startups that raised less than $5M can apply for $1,000 of credit.1 |
| 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. | The v3 REST API, a Python SDK (honcho-ai), a TypeScript SDK (@honcho-ai/sdk) and a CLI.2 |
Pricing
Both price learning by the token and questions by reasoning level, with the same five levels, and both include context. Recall bills each message as at least 50 tokens.
| What you pay for | Recall | Honcho1 |
|---|---|---|
| Learning from messages, per 1M tokens | $1.75 | $2.00 |
A question at minimal | $0.0008 | $0.001 |
A question at low | $0.005 | $0.01 |
A question at medium | $0.025 | $0.05 |
A question at high | $0.05 | $0.10 |
A question at max | $0.25 | $0.50 |
| Context for a prompt | Included | Included |
| Background consolidation | Included (the night shift) | Included (dreaming) |
| Free credit | $5 once | $5, then $5 a month with a card on file |
A worked example. A support assistant that learns from 1,000,000 message tokens a month and asks 500 questions at low pays $4.25 on Recall ($1.75 + $2.50) and $7.00 on Honcho at its list prices ($2.00 + $5.00), if Honcho counts the same tokens. Honcho's monthly $5 of credit, with a card on file, covers most of that bill; Recall's $5 is given once.
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 | Honcho | How each was graded | Compare directly? |
|---|---|---|---|---|
| LongMemEval-S | 91.0% (455 of 500) | 90.4% (500 questions)4 | Recall: LongMemEval's official judge, gpt-4o-2024-08-06 with the benchmark's own prompts. Honcho: GPT-4o with the paper's prompt; answers by Claude Haiku 4.5, memory built by gemini-2.5-flash-lite.4 | The closest of any pair on these pages: the same questions and a GPT-4o judge. The answer models differ, and both figures are self-reported. |
| BEAM 100K | 0.683 (all 400 questions) | 0.630 (400 questions)5 | Recall: BEAM's own evaluation code and judge prompts, graded by gpt-4.1-mini as BEAM's code does. Honcho: graded per the BEAM paper, judge model not named; answers by Claude Haiku 4.5.5 | Roughly: both use BEAM's scoring, but Honcho does not name its judge model. |
| BEAM 500K | 0.670 (all 700 questions) | 0.6465 | As for BEAM 100K. Plastic Labs' blog gives 0.649.4 | Roughly, as for BEAM 100K. |
| BEAM 1M | Not run yet | 0.6185 | As for BEAM 100K. | Recall has no figure to compare. |
| BEAM 10M | Not run yet | 0.4095 | As for BEAM 100K. | Recall has no figure to compare. |
| LoCoMo | Not run in full. On a 75-question sample: 60.0% with a strict judge, 79.2% with a lenient one | 89.9% (1,540 questions)5 | Honcho: judged by an LLM, model not named.5 | No: Recall has no full LoCoMo run. |
Honcho also reports 92.6% on LongMemEval-S with Gemini 3 Pro, without the full setup on the page.5 Plastic Labs publishes its configuration and raw results on GitHub, linked from its evals page.5
Hosting and licences
Hosted. Both run a hosted service. Recall stores data in Frankfurt (EU); its models run through providers in the United States (see the data processing agreement).
Self-hosted. Honcho's server is open source under the AGPL-3.0, an OSI licence.2 Recall's server is source-available under the Functional Source License: you may run it in production for your own products and change it, but not offer it to others as a competing memory service, and each version becomes Apache 2.0 two years after release. Honcho needs Postgres with pgvector and refuses to start without an LLM provider configured.3 Recall runs as one process with Postgres and pgvector or an embedded database, and without a provider key it runs on an offline mock model, which is enough to try the API.
API
Recall implements Honcho's v3 REST API: the same paths, bodies, errors, filters and pagination, and Honcho's MCP tool names. We ran Honcho's own published clients, honcho-ai and @honcho-ai/sdk version 2.5.1, against Recall with only the base URL and the API key changed, and a whole app worked; the TypeScript run is a test in Recall's repository that runs on every change. The differences that matter (reads that do not create, different models, bounded peer cards, retried webhooks, self-hosting settings) are in Switch from Honcho.
When to choose Honcho
- You want an open-source licence. Honcho's server is AGPL-3.0; Recall's server is source-available and may not be offered to others as a competing memory service.
- You need results at very long histories. Honcho publishes BEAM figures at 1M and 10M tokens; Recall has not run those yet.
- Your usage is small and steady. Honcho's $5 of credit every month, with a card on file, can cover it; Recall's free credit is given once.
- You are a startup that qualifies for Honcho's $1,000 credit programme, or you want Plastic Labs' own reasoning model.
When Recall fits better
- You want lower list prices: Recall charges less for learning ($1.75 against $2.00 per 1M tokens) and for a question at every level.
- You want answers you can check: chat can return citations, a confidence, and an explicit "I don't know", and facts carry the dates they were true.
- You want your data stored in the EU.
- Either way, trying the other is a change of base URL and key.
See all comparisons, or try Recall with $5 of free credit: quickstart.
Sources
Read on 13 October 2026.