Playground

Playground

Watch Mnemosyne think, calculate, and recall — entirely in your browser. No backend. Drag, click, mouse around.

ROI calculator

What does Mnemosyne actually save you per month? Move the sliders to your numbers and watch.

$440without Mnemosyne
$297with Mnemosyne
$143saved / month
57.2Mtokens saved / month

Live recall pipeline

A query enters at the left and propagates through the 7-stage pipeline. Hover any stage for what it does. Hit “Run again” to replay.

Recall pipeline
query: "coffee preferences"
step 1
Auth + RLS
0.4ms
step 2
Query normalize
0.9ms
step 3
Hot cache
1.8ms
step 4
Warm pointer + LSH
4.2ms
step 5
FTS prefilter
6ms
step 6
HNSW vector
9.5ms
step 7
Rerank + compose
13.8ms
Hover a stage to see what it does. Total path: ~14ms p50 cold, <2ms p50 hot.

Memory graph

A live force simulation of a tiny multi-typed memory. Every dot is a typed node — fact (violet), entity (cyan), decision (amber). Mouse around.

● fact● entity● decisionlive force simulation

Latency by tier

Lower bars are faster. We highlight Mnemosyne’s hot tier in violet — that’s the p50 your agent feels for already-seen queries.

Cold (p99)
195ms
Cold (p95)
140ms
Cold (p50)
65ms
Warm (p95)
14ms
Warm (p50)
8ms
Hot (p95)
4ms
Hot (p50)
2ms

Compare quality

F1 score on LongMemEval — higher is better. We didn’t pick the easiest benchmark — this is the hardest public one for agentic memory.

Letta (MemGPT)
0.778
Mem0
0.775
Cognee
0.806
Zep
0.824
Mnemosyne
0.854

Try the real thing

Convinced? Take it for a spin in 60 seconds:

docker run --rm -d \
  -e DATABASE_URL=postgres://postgres:mnemo@host.docker.internal:5432/mnemo \
  -e MNEMO_KEY=mns_live_demo \
  -p 3000:3000 ghcr.io/lucasmailland/mnemosyne-server:latest
 
curl http://localhost:3000/v1/health

Then the full Quickstart walks you through writing facts and recalling them with real embeddings.