On 2026-11-22, Mem0 had the most GitHub stars of the 8 agent memory projects we checked (69,350). Hindsight was second (62,669). Mem0’s mem0ai package also had the most downloads, 2,446,544 on PyPI and 1,045,328 on npm last month. These memory library statistics were collected by aiagentmemory.org from the GitHub REST API, pypistats.org and the npm registry on 2026-11-22 at 22:17 UTC. We did not run any of these projects ourselves.
For the same kind of numbers on other tools, see agent memory statistics, graph database statistics and local LLM statistics. For a side-by-side reading of two of these projects, see Mem0 vs Hindsight.
Memory library GitHub and download statistics
Table 1. GitHub and download statistics for 8 agent memory projects, collected 2026-11-22. Sorted by stars. The PyPI and npm columns are last-month downloads for the main package each project publishes; see the notes below the table.
| # | Project | Repository | Stars | Forks | Created | Last push | License | PyPI/month | npm/month |
|---|---|---|---|---|---|---|---|---|---|
| 1 | Mem0 | mem0ai/mem0 | 69,350 | 8,247 | 2023-06-20 | 2026-11-18 | Apache-2.0 | 2,446,544 | 1,045,328 |
| 2 | Hindsight | vectorize-io/hindsight | 62,669 | 7,400 | 2025-10-30 | 2026-11-22 | MIT | 1,044,982 | 260,092 |
| 3 | Cognee | topoteretes/cognee | 33,472 | 3,544 | 2023-08-16 | 2026-11-20 | Apache-2.0 | 178,079 | - |
| 4 | Graphiti | getzep/graphiti | 33,004 | 3,402 | 2024-08-08 | 2026-11-20 | Apache-2.0 | 896,334 | - |
| 5 | Supermemory | supermemoryai/supermemory | 31,689 | 2,781 | 2024-02-27 | 2026-11-22 | MIT | 159,284 | 850,698 |
| 6 | Letta | letta-ai/letta | 25,513 | 2,687 | 2023-10-11 | 2026-10-18 | Apache-2.0 | 127,230 | 411,243 |
| 7 | Honcho | plastic-labs/honcho | 8,237 | 1,003 | 2023-09-10 | 2026-11-21 | AGPL-3.0 | 282,067 | 135,706 |
| 8 | LangMem | langchain-ai/langmem | 1,786 | 210 | 2025-01-21 | 2026-11-16 | MIT | 837,337 | - |
The packages counted are: mem0ai (PyPI and npm), graphiti-core (PyPI), cognee (PyPI), letta-client (PyPI) and @letta-ai/letta-client (npm), honcho-ai (PyPI) and @honcho-ai/sdk (npm), supermemory (PyPI and npm), hindsight-client (PyPI) and @vectorize-io/hindsight-client (npm), and langmem (PyPI). “-” means our list has no npm package for that project, so there is no npm count.
What the numbers show
- Mem0 leads on stars and on downloads. It has the most stars (69,350) and the most PyPI and npm downloads of the 8 projects.
- Hindsight is second on stars and on PyPI downloads. Its repository was created on 2025-10-30, so 62,669 stars came in about 13 months. Its PyPI client had 1,044,982 downloads last month, close to half of Mem0’s count.
- Supermemory’s npm count is high. Its
supermemorynpm package had 850,698 downloads, second to Mem0’s npm package. Its server source is not public, so you cannot self-host it from the repository. - Letta’s last push was 2026-10-18. It is still listed as active on this table, but it has not pushed for about five weeks as of this collection.
- Download counts are not like for like. Mem0’s
mem0aiand Cognee’scogneeare full libraries. Hindsight, Letta and Honcho count client or SDK packages. Compare the trend within a row more than the numbers across rows.
Download counts include CI runs, container builds and packages installed as dependencies, so they are not user counts. Monthly counts also swing from one month to the next. Read a change of a few tens of percent as a signal, not a trend.
Memory: where Hindsight fits
Hindsight is an open-source memory server, made by Vectorize, under the MIT license. The full server is in the open-source version, so you can self-host it without a paid tier. Its numbers below come from its README and from Mem0’s own comparison page.
- Self-hosting is free. One
docker runstarts the server with a built-in Postgres saved to a volume. Storage is PostgreSQL with pgvector. - Long histories. On Mem0’s head-to-head page, Hindsight scored 64.1 on BEAM 10M against Mem0’s 48.6, and 73.9 against 64.1 on BEAM 1M. Both sets of figures come from Mem0’s comparison page, not from an independent test.
- The limit: writes call an LLM. Retain uses a model to extract facts, time, entities and relationships, so every write costs model calls. Hindsight also returns more context per retrieval than Mem0 by default, so set
max_tokenslower if that matters to you.
Methodology
When. All numbers were collected on 2026-11-22, starting at 22:17 UTC, by aiagentmemory.org with the same script used for agent memory statistics.
Which APIs.
- GitHub:
GET https://api.github.com/repos/{owner}/{repo}. Fields used:stargazers_count,forks_count,created_at,pushed_at,license.spdx_id. - PyPI:
GET https://pypistats.org/api/packages/{package}/recent, last-month total. - npm:
GET https://api.npmjs.org/downloads/point/last-month/{package}, last-month total.
Which packages. Each package name was checked on PyPI or the npm registry against the project’s own repository before the run. What we did not count. Zep’s platform and its clients are not in this table. Memory projects that ship an MCP server are in MCP memory statistics, which counts a different kind of package.