The cognition role (memory.extractionModel, ADR-23) lets a small local model own memory-writing while chat uses anything. The shelf recommends LFM2.5-2.6B for it. Liquid also publishes task-specific "Nanos", including LiquidAI/LFM2-1.2B-Extract, tuned for structured extraction, which is exactly the extractor's job.
The experiment: run the extraction prompt in src/memory-extract.js against LFM2-1.2B-Extract, LFM2.5-2.6B, and Qwen3 4B on the same twenty exchanges; report precision (facts that are true and durable), recall (facts a person would have kept), format compliance (does it follow the bullet contract), and speed on CPU. If Extract wins, the shelf entry changes and the guide gets a paragraph; if it does not, the report is still worth having.
Done when: the comparison is posted here with the exchanges used (synthetic, no real data).
The cognition role (
memory.extractionModel, ADR-23) lets a small local model own memory-writing while chat uses anything. The shelf recommends LFM2.5-2.6B for it. Liquid also publishes task-specific "Nanos", includingLiquidAI/LFM2-1.2B-Extract, tuned for structured extraction, which is exactly the extractor's job.The experiment: run the extraction prompt in
src/memory-extract.jsagainst LFM2-1.2B-Extract, LFM2.5-2.6B, and Qwen3 4B on the same twenty exchanges; report precision (facts that are true and durable), recall (facts a person would have kept), format compliance (does it follow the bullet contract), and speed on CPU. If Extract wins, the shelf entry changes and the guide gets a paragraph; if it does not, the report is still worth having.Done when: the comparison is posted here with the exchanges used (synthetic, no real data).