This version is approximately 2,900 characters—well under your 4,000-character limit—while retaining all the technical authority and key quotes from the original.
Backboard.io Sets New State-of-the-Art Benchmarks for AI Memory
Backboard.io today announced record-breaking results across two leading AI memory benchmarks, LoCoMo and LongMemEval, solidifying its position as the foundational infrastructure for production-grade agentic systems.
Breakthrough Performance
An independent evaluation by NewMathData (AWS Small Partner of the Year) measured Backboard on the LongMemEval benchmark. Backboard achieved 93.4% overall accuracy, the highest publicly reported result to date.
During review, evaluators noted that Backboard’s score is likely a "conservative lower bound." In several instances, Backboard provided more precise, semantically accurate answers than the benchmark’s "gold" standard by incorporating factual context from the broader interaction.
These results follow Backboard’s 90.1% accuracy on the LoCoMo benchmark. While most systems sacrifice long-horizon persistence for short-term precision, Backboard excels at both.
Redefining "Memory" in AI
In an analysis for the Ottawa Business Journal, Adyasha Maharana, creator of LoCoMo and research scientist at Databricks, noted that breaking the 90% threshold requires "superhuman consistency."
"Strictly speaking, [feeding a full conversation as a prompt] is not memory," Maharana explained. "The system built by Backboard.io is a far better attempt at simulating memory as it manifests in humans. It is practical, cheaper, scalable, and doesn’t rely solely on brute-force LLM processing."
A Unified AI Stack
Backboard.io is not a plugin or a wrapper; it is a unified infrastructure stack. By treating memory as first-class infrastructure, Backboard eliminates the need for enterprises to stitch together fragile open-source components.
The platform provides a single API for:
Persistent long-term and shared agent memory
Native embeddings, vectorization, and RAG
Access to 17,000+ LLMs (including BYO-key options)
This architecture allows models to be swapped and agents to coordinate without losing continuity or rewriting retrieval strategies.
Making Agentic AI Practical
"Agentic AI becomes meaningful when agents can remember, coordinate, and operate over time," says Rob Imbeault, founder of Backboard.io. "Solving memory is the prerequisite."
Backboard’s Active Temporal Resonance framework preserves meaning as interactions unfold, ensuring systems remain consistent and auditable at scale. This focus on durability stems from Imbeault’s experience founding Assent, a platform trusted by Fortune 100 companies for complex compliance.
What’s Next: Introducing Switchboard
With its memory foundation validated, Backboard will soon launch Switchboard—a new capability helping developers evaluate how different AI configurations behave under real-world constraints.
"We didn’t build Backboard to chase benchmarks," said Imbeault. "We built it to solve real problems... the benchmarks just happened to confirm what we were already seeing in practice."
> This version is approximately 2,900 characters—well under your 4,000-character limit—while retaining all the technical authority and key quotes from the original.
This version is approximately 2,900 characters—well under your 4,000-character limit—while retaining all the technical authority and key quotes from the original.
Backboard.io Sets New State-of-the-Art Benchmarks for AI Memory Backboard.io today announced record-breaking results across two leading AI memory benchmarks, LoCoMo and LongMemEval, solidifying its position as the foundational infrastructure for production-grade agentic systems.
Breakthrough Performance An independent evaluation by NewMathData (AWS Small Partner of the Year) measured Backboard on the LongMemEval benchmark. Backboard achieved 93.4% overall accuracy, the highest publicly reported result to date.
During review, evaluators noted that Backboard’s score is likely a "conservative lower bound." In several instances, Backboard provided more precise, semantically accurate answers than the benchmark’s "gold" standard by incorporating factual context from the broader interaction.
These results follow Backboard’s 90.1% accuracy on the LoCoMo benchmark. While most systems sacrifice long-horizon persistence for short-term precision, Backboard excels at both.
Redefining "Memory" in AI In an analysis for the Ottawa Business Journal, Adyasha Maharana, creator of LoCoMo and research scientist at Databricks, noted that breaking the 90% threshold requires "superhuman consistency."
"Strictly speaking, [feeding a full conversation as a prompt] is not memory," Maharana explained. "The system built by Backboard.io is a far better attempt at simulating memory as it manifests in humans. It is practical, cheaper, scalable, and doesn’t rely solely on brute-force LLM processing."
A Unified AI Stack Backboard.io is not a plugin or a wrapper; it is a unified infrastructure stack. By treating memory as first-class infrastructure, Backboard eliminates the need for enterprises to stitch together fragile open-source components.
The platform provides a single API for:
Persistent long-term and shared agent memory
Native embeddings, vectorization, and RAG
Access to 17,000+ LLMs (including BYO-key options)
This architecture allows models to be swapped and agents to coordinate without losing continuity or rewriting retrieval strategies.
Making Agentic AI Practical "Agentic AI becomes meaningful when agents can remember, coordinate, and operate over time," says Rob Imbeault, founder of Backboard.io. "Solving memory is the prerequisite."
Backboard’s Active Temporal Resonance framework preserves meaning as interactions unfold, ensuring systems remain consistent and auditable at scale. This focus on durability stems from Imbeault’s experience founding Assent, a platform trusted by Fortune 100 companies for complex compliance.
What’s Next: Introducing Switchboard With its memory foundation validated, Backboard will soon launch Switchboard—a new capability helping developers evaluate how different AI configurations behave under real-world constraints.
"We didn’t build Backboard to chase benchmarks," said Imbeault. "We built it to solve real problems... the benchmarks just happened to confirm what we were already seeing in practice."
> This version is approximately 2,900 characters—well under your 4,000-character limit—while retaining all the technical authority and key quotes from the original.
lmfao