Selection over volume
Does a short, relevant memory set outperform a much larger context full of weakly related material?
PROJECT MNEMOSYNEACTIVE RESEARCH / 2026
MEMORY FOR SMALL LANGUAGE MODELS
What if the
model did not
have to hold
everything?
Mnemosyne investigates the space between model capacity and the information system around it.
MNEMOSYNE / MEMORY LABCONFIGURATION / SELECTED
EXTERNAL MEMORY4 / 6 MEMORY ITEMS AVAILABLE TO THE MODEL
✓Prior decision
✓Task constraint
✓Working example
✓Domain fact
—Old transcript
—Loose document
SMALL LANGUAGE MODEL
SLMCONTEXT IN
A smaller, task-specific context.
WHAT WE MEASURE
QUALITY?
LATENCY?
COST?
Does selection help?
INPUT / MODEL + MEMORY + TASKTHIS IS A RESEARCH CONFIGURATION — NOT A PERFORMANCE CLAIMOUTPUT / EVIDENCE
THE RESEARCH BET
Better architecture can
sometimes buy back scale.
A small language model cannot become a large model by being given more text. But the right facts, previous decisions, examples and constraints may change the task it is being asked to solve. We are measuring exactly where that helps.
CORE QUESTION
How much capability
can relevant memory
recover?
Not “does RAG work?” A more specific question: what is the trade-off between internal parameter capacity and external contextual capacity for a real task?
THE STUDY DESIGN
BASELINE → MEMORY CONDITIONS → COMPARISONCapability boundaries
Which tasks improve through context, and which still require capability that only model scale provides?
Useful efficiency
Where can a smaller system retain enough quality to make the trade-off in cost and latency worthwhile?
EXPERIMENTAL PROTOCOL
Same task.
Different system.
Prompt + parameters only.
Change selection, structure and amount of context.
Compare quality, cost and response time on the same work.
WHAT WE WILL NOT CLAIM
Every useful result
has an edge.
Mnemosyne is not built to declare that small models replace larger ones. The useful outcome is a map of the conditions where memory helps, where it fails, and what the system costs when it does.
See the rest of the lab ↗Interested in the research?
Talk to the Mnemosyne team. ↗