Memory
Context thatcarries forward.
Sofie remembers your role, your products, your preferences and the decisions your team has made, as governed memory linked to its evidence. Work starts where it left off instead of from a blank prompt.
LB-204
Compound
Fed-batch upstream
Process
Protein A capture
Process
IL-23p19
Gene/Protein
Plaque psoriasis
Disease
FDA
Regulatory Body
Northwind · Alameda
Site
Northwind · Durham
Site
Observations
LB-204 hold limits
94%Clarified harvest validated to 72 h at 2–8 °C; HT-7 has no supporting study yet.
3 sources
Northwind engineering run
88%SUB-2000 qualified; engineering run held for the week of Nov 10.
2 sources
Rules
Format outputs for GMP documentation
Cite page numbers for every workspace claim
Use ICH Q5E terminology for comparability
Everynewchatwithagenericassistantstartsfromzero.Sofiekeepswhatmatters,checksitagainstitssourcebeforesavingit,andletsyouradministratorsdecidewhatitmaykeep.
Capabilities
What Memory does
01
Learns as you work
Facts are extracted from conversations automatically and consolidated each night, so memory stays tidy as it grows.
02
One summary per topic
Observations keep a single maintained summary for each topic, linked to its evidence. Every write is source-checked, and different people are never blended together.
03
A biopharma knowledge graph
Twelve entity types, including compounds, diseases, genes and proteins, regulatory bodies and processes, connected as Sofie learns.
04
Hybrid retrieval
Semantic, keyword, graph and recency signals are merged into one ranking, so the right memory surfaces for the question at hand.
05
Confidence and contradictions
Each memory carries a confidence score, and Sofie detects when something new contradicts what it already knows.
06
Rules and continuity
Directives are standing rules, such as your deviation numbering. Episodes carry the thread of a work session into the next one.
In every answer
Sofie already knows the program.
Ask about LB-204 and Sofie knows it is a monoclonal antibody moving to a new receiving site, which hold times were in question last week, and that your QA lead wants ICH Q9(R1) language. You spend the conversation on the problem, not the background.
- Semantic, keyword, graph and recency signals in one ranking
- Memories linked to the evidence they came from
- Rules applied without being restated
HT-7 hold-time support
Deep researchSources
6 of 18 citedHTS-014 Hold-Time Study Report.pdf
p. 14 · exact
21 CFR 211.111 · Time limitations on production
ecfr.gov
Process Validation: General Principles and Practices
fda.gov · Guidance
LB-204 Transfer Protocol v3.docx
§6.2 · exact
BR-2291 Executed Batch Record.pdf
p. 22 · exact
Designing hold-time studies for biologics intermediates
bioprocessintl.com
Governance
Your administrators decide what Sofie keeps.
Memory can be enabled or disabled for a user, a team or the whole site. It never crosses organization boundaries and it respects your retention window. The Memory Intelligence page shows what Sofie knows in an overview, a knowledge graph, a timeline, a library and your rules.
- Controls at user, team and site level
- Never shared across organizations
- Follows your retention window
- SiteLarkspur Biologics
- TeamQuality Assurance
- TeamBusiness Development
- UserJ. Okafor
Boundary
This organization only
Retention
Follows your retention window
Rule
Use our deviation numbering (DEV-YYYY-NNN) in every draft.
00:00:00:00
Product film · Context that carries forward · In production
Memory
- on the LongMemEval benchmark (455 of 500), fresh full run
- 91.0%
- entity types in the knowledge graph
- 12
- retrieval signals merged into one ranking
- 4
Questions
Frequently asked
Who controls what Sofie remembers?
Your administrators. Memory can be enabled or disabled at the user, team or site level, and it respects your organization's retention window.
Can memory leak between organizations?
No. Memory never crosses organization boundaries, and each customer runs in its own isolated deployment.
How does Sofie avoid remembering the wrong thing?
Every write is checked against its source, facts about different people are never blended, and each memory carries a confidence score. Contradictions are detected rather than silently overwritten.
What is LongMemEval?
A public benchmark for long-term conversational memory. Sofie scored 91.0%, 455 of 500 questions, on a fresh full run.
Works with
Part of one connected platform
Research & Chat
Deep research across your files and the web, every claim cited.
Workspaces
Every program, site and client, with its files, people and history in one place.
Orchestrations
Specialized agents coordinated across multi-step work, with human review points.
Sofie Voice
Talk through the work while Sofie builds it beside you in split screen.
Start with one workflow.
Measure it. Then expand.
In a working session we choose a priority workflow, define your standards and review points, and agree how to measure speed, quality and capacity. Then you evaluate Sofie's work before deciding whether to expand.