Tech Reports
Read cognee's tech reports: how we evaluate agent memory on public benchmarks such as BEAM and LoCoMo, what the Rust engine does on device, and what persistent memory costs in tokens. Each post covers the method and the results, and links to the full report.
Latest
cognee on LoCoMo: 91% on Long-Term Conversational Memory
cognee answers 91% of 1,762 evaluated LoCoMo questions correctly with standard memory components, one answer prompt and a strict LLM judge. Exclusions, prompts and records are published.
cognee on BEAM: SOTA Results Without a Benchmark-Specific Memory System
cognee beat SOTA on BEAM's 100k-token setting by 6.5% and matched SOTA at 10M tokens using only default open-source features — no custom benchmark-specific architecture.
Technical Note: Understanding the Token Cost of Persistent AI Memory
Persistent memory trades an upfront ingestion cost for cheaper queries. We measure where the tokens go in cognee, model the trade-off, and find break-even at roughly 23–26 repeated queries — after which the gap keeps widening.
cognee on-device: Bringing Agent Memory to the Edge
The cognee core rebuilt in Rust, bringing the full agent memory pipeline — graph construction, embeddings, and retrieval — to phones, robots, and offline environments without a server-side stack.
cognee on LoCoMo: 91% on Long-Term Conversational Memory
cognee answers 91% of 1,762 evaluated LoCoMo questions correctly with standard memory components, one answer prompt and a strict LLM judge. Exclusions, prompts and records are published.
cognee on BEAM: SOTA Results Without a Benchmark-Specific Memory System
cognee beat SOTA on BEAM's 100k-token setting by 6.5% and matched SOTA at 10M tokens using only default open-source features — no custom benchmark-specific architecture.
Technical Note: Understanding the Token Cost of Persistent AI Memory
Persistent memory trades an upfront ingestion cost for cheaper queries. We measure where the tokens go in cognee, model the trade-off, and find break-even at roughly 23–26 repeated queries — after which the gap keeps widening.
cognee on-device: Bringing Agent Memory to the Edge
The cognee core rebuilt in Rust, bringing the full agent memory pipeline — graph construction, embeddings, and retrieval — to phones, robots, and offline environments without a server-side stack.
Coding Agent Memory with cognee: Hooks, Capture, and Recall
Keenable × cognee: Giving Web Retrieval Persistent Memory
