1. Compare on one test
Overall rank comes from one independent run. Vendor claims and other benches stay in their own tabs.
For builders who ship AI agents
Your AI isn't dumb. It has amnesia.
Every new chat, your agent forgets the work. You explain it again. Memory tools fix that. Every vendor says they are #1. We only rank tools that share the same test, so you pick the right one in minutes, not weeks.
2. Pick the fit
See how each tool stores memory, where it runs, and who it is built for. Then open the docs.
3. Stay current
Amnesia Brief is one short email a week: what changed, which claims moved, and the scoreboard.
01 / Directory
Nine memory tools worth a look
These help your agent remember across sessions. Summaries come from public docs. Nobody paid for a spot.
| Name | One-liner | Memory model | Deploy | Best for | Link |
|---|
No tools match that filter.
02 / Compare
Same test. Then rank.
Pick a benchmark first. Only tools on that test can beat each other. Overall public rank uses one independent run. Nothing else.
Watch the gap. Vendors post 74% to 95% on LoCoMo. One independent Sep 2026 run put the same tools between 37% and 58% on a single conversation. Setup moves the score more than the brand does.
No rows match that filter on this benchmark.
03 / How we rank
Apples to apples, or no rank
If two numbers came from different tests, they do not share a rank. That rule keeps marketing out of the #1 slot.
What sets overall rank
Only the Independent Sep 2026 LoCoMo conversation 26 run by Hamza Shabbir. Same chat. Same answering model (gpt-4o-mini). Same grading. Tools that skipped that test do not get an overall rank.
- Metric: recall on retrieved text (LLM judge).
- Tie-break 1: higher score wins.
- Tie-break 2: lower search latency (p50) wins.
- Tie-break 3: earlier published date wins.
Source: Hamza Shabbir benchmark ↗
What does not set overall rank
- Vendor LoCoMo claims (their method, their grader).
- LongMemEval scores (a different benchmark).
- Competitor papers that rerun each other.
- BEAM or other benches until several tools share one public run.
- Mixing any of the above into one #1 list.
Those tabs stay up so you can read the claims. They only sort inside their own list.
04 / Newsletter
Amnesia Brief
One email a week. What changed in AI memory, which claims moved, and the monthly scoreboard. Three minutes. No spam.
05 / About
Nobody pays for a ranking
We compare tools first. Then we test them ourselves. If we sell a featured spot later, it will say Featured in plain sight.
What we believe
Smart without memory is just a brainstorm. Memory lets your agent keep your decisions, preferences, and lessons. You and your team stop starting from zero.
What we will not do
Hide paid spots in a best-of list. Invent categories to sell links. Show a vendor number without its source and a vendor label. Trust is the product.
06 / Contact
Send a message
Wrong number? Tool we missed? Want something tested? Write us. A real person reads every message.