DAKERAAgent memory is one of those things that's easy to explain but surprisingly hard to evaluate — until you run it yourself. The Dakera Playground is a live API demo that lets you do exactly that, from your browser, in under a minute.
What the playground is
The playground connects directly to a shared Dakera sandbox running our production server. When you click Run on any scenario, you're sending a real API request — not a mock. The response comes back from the same memory engine that backs production deployments, with the same importance-based recall, the same temporal weighting, and the same entity extraction.
The rate limit is 30 requests per minute with 50 memory slots per session. That's enough to run every scenario, experiment with parameters, and see the full range of Dakera's recall behavior — all without creating an account or configuring anything.
Seven scenarios, each showing a distinct capability
Each scenario uses a realistic AI agent persona — a fintech advisor, a clinical assistant, an enterprise knowledge bot — so the code isn't toy data. Here's what you can explore:
Five languages, one API
Every scenario runs in Python, JavaScript, Rust, Go, and cURL. The code editor is live — you can edit the example, adjust parameters with the importance slider and top_k control, then run your modified version directly against the sandbox. All five SDKs hit the same REST endpoints.
Three modes beyond the guided tour
Compare Mode runs the same recall query at top_k=10, top_k=5, and top_k=3 simultaneously — letting you see exactly how narrowing the retrieval window changes what your agent knows. It's a fast way to calibrate the breadth vs. precision tradeoff for your use case.
API Explorer gives you a freeform request editor against six live endpoints: /memory/store, /memory/recall, /memory/search, /agent/memories, /memory/importance, and /knowledge/query. This is the closest thing to having a Dakera instance in front of you without running one locally.
What this shows about Dakera's recall model
Every recall response includes a smart score — a composite of vector similarity, importance weight, and temporal recency. That's not a single embedding lookup; it's a pipeline. The semantic recall scenario makes this concrete: the query "concurrent resource access" lands near "Redis mutexes" in embedding space because they describe the same engineering concept, not the same words. The importance weight means a high-priority memory stays near the top even as time passes. And the temporal factor means recent memories get a recency boost when the query implies current context.
Together, these three signals produce the 88.2% accuracy on the LoCoMo long-context memory benchmark — a dataset of 30-turn conversations where the agent must retrieve specific facts mentioned much earlier. No other open-core memory server has published a score on this benchmark.
Benchmark context: LoCoMo Cat3 tests temporal inference — answering questions that require reasoning about when events happened, not just what happened. Dakera scores 68.5% on Cat3, making it the hardest category. The playground's Time-Aware scenario demonstrates exactly this class of query.
From playground to production
The playground API endpoint (dakera.ai/api/playground/v1) uses the same HTTP API as a self-hosted Dakera instance. Swapping in your own server URL and API key is the only change needed to take playground code to production. The SDK quickstarts walk through that setup for all four languages — it typically takes under five minutes from a cold start.
Ready to try it?
Seven live scenarios. Five languages. No account needed. Open the playground now or get started with the quickstart guide.