Memory for agents
Vendors publish 90%+ on LoCoMo and LongMemEval. The one independent comparison we found graded five systems on the text they retrieved from the same 419-turn conversation, and the best scored 57.6%. A large context window alone scores 95.6% on LongMemEval-S, which makes this category as much about token cost as capability.
Catalog only, not yet measuredGitHub, via our catalog · catalog compiled Nov 7, 2026
Adoption: GitHub stars
Independent tests, beside a vendor claimhamzashabbir.dev five-system benchmark (published harness); LongMemEval-S · published 2026 (month not stated)
Vendor claims against independent tests
92.5%
Mem0 on LoCoMo
57.6%
best independent score (GoodMem)
37.0%
lowest independent (Zep/Graphiti)
95.6%
a large context alone, LongMemEval-S
Vendor-reported / commissionedIndependent benchmark
Tools
By GitHub stars, most first; tools without a public repo last, alphabetically. Links go to each vendor's own site.
| Tool | What it does | Evidence | Price or terms |
|---|---|---|---|
| Mem0 | managed vector + graph memory | 68.6k GitHub stars | free (own LLM key) + custom |
| Cognee | graph + vector memory, Postgres-only at 1.0 | 33.1k GitHub stars | open source; cloud paid |
| Zep (Graphiti) | temporal knowledge-graph memory | 32.6k GitHub stars (Graphiti repo) | free 10k credits; Flex $125/mo |
| Letta | stateful agents; agent-managed memory | 25.4k GitHub stars | per active agent, no base fee |
| LangMem | memory SDK for LangGraph | 1.8k GitHub stars | free, open source |
Caveats
- LoCoMo is close to saturated: dumping the whole conversation into the prompt now scores competitively.
- The open Agent Memory Benchmark (AMB) is run by a memory vendor (Hindsight's maker), though its data and scoring are public.