AI agents do not fail because they lack memory. They fail because their context does not evolve. We spent the past few months on this problem. Today Maximem Synap is live. - 90.2% on LongMemEval. - 15ms P50. - 10 frameworks on day one.

The numbers: Maximem Synap: 90.2% Next closest: 71.3% Then: 63.8% and 57.5% Same harness, same hardware, same evaluation prompts. P50: 15ms. P99: under 300ms. SDK is open source. Eval harness is open source. Free tier is live.

The deep-dives, if you want them: - Why we built Synap: - Benchmark results: - Technical architecture: First 100 customers get 3 months Pro ($500/mo) free (no credit card needed)

@IT_Kabootar interesting man! Happy to connect

@IT_Kabootar Super cool! Congrats

@IT_Kabootar context that evolves is a real unlock the next layer is making sure the actions that follow are just as reliable deterministic execution, settled outcomes, verifiable logs memory + execution is the full stack.

This is super interesting. I’ve been working pretty closely on memory and context systems for agents, so the 15ms P50 really caught my eye. It almost feels too good to be true 😄 Would love to understand this better. Is the 15ms measured with or without LLM calls in the loop? Is this latency local or over a network, and under what conditions? Does this include only cache hits or also cold retrievals? And do you have any numbers for P95 or P99 under real-world load? Really curious how you’re achieving this. Impressive if it holds up.

15ms is the agent experienced time :)

@IT_Kabootar This is so cool!