PRISM preserves your original project files and builds searchable, traceable context for AI. Sources, versions and evidence stay connected as your projects and models change.
Project files feed PRISM, which preserves original sources and builds a searchable Project Brain. An AI agent receives evidence with its source, version and relevant passage. The model reasons from that evidence.
Keep the originals.Project history stays available.
Trace the evidence.Retrieved evidence links back to its source.
Change the model.Your project context stays yours.
Context that holds up
When sources disagree, keep both in view.
Projects change. Notes conflict. Older decisions still matter. Useful AI needs that history, including the moments when there isn’t enough evidence to be certain.
PRISM keeps the source and version with the result, so the model can explain a disagreement rather than hide it.
Example · Conflicting project records
“Has this change been approved?”
Runbook v4.2
“The change is approved.”
Source passage · Section 5
Known issues v1.3
“Approval remains on hold.”
Source passage · Issue 12
Possible model interpretation
These records disagree. Approval cannot be established from this evidence alone.
Both sources remain visible. A search score cannot settle the disagreement.
How PRISM works
Preserve first. Derive second. Reason last.
Original sources are durable. The Project Brain is derived and rebuildable. The model receives evidence and remains responsible for interpreting it.
PRISM preserves sources, builds a derived Project Brain and supplies read-only evidence search. Model reasoning is a separate responsibility.
PRISM · Evidence infrastructure
01
Project Sources
Original files Identity and versions Durable history
Preserved
02
Project Brain
Structured material Lexical + semantic indexes Links to the source
Derived · Rebuildable
03
Evidence search
Relevant passages Sources and versions Retrieval facts
Read-only
Model reasoning
04
Your AI model
Compare sources Explain uncertainty Reason and cite
Interpretation
Source storage and retrieval are designed to run locally. Evidence sent to a cloud model leaves that local environment.
A closer look at retrieval+
RAG is part of the approach.
Retrieval-augmented generation helps a model work with relevant material. PRISM adds durable source history, project boundaries and evidence handling around that retrieval.
Find relevant material
Lexical and semantic search can work together. The system reports the retrieval mode it actually used.
Keep the project boundary
Search scope comes from the active trusted project, rather than an arbitrary path supplied by a model.
Separate search from change
Evidence search is read-only. Source-to-Brain reconciliation is a separate operation.
Keep scores in perspective
Similarity and ranking describe retrieval. They do not establish that a statement is true.
In development at Comans
Practical foundations for project intelligence.
We’re developing and testing PRISM inside Comans, starting with preserving information and making evidence easier to find and use.
Preserve project history
Keep original files, logical source identities and versions together.
Work across formats
Prepare documents, spreadsheets, email, images and archives for AI workflows.
Find words and meaning
Combine lexical and semantic retrieval without depending on embeddings alone.
Keep evidence connected
Return source, version and relevant passage information with each result.
Read without rewriting
Search within the active project, with changes handled separately.
Show what actually ran
Report the retrieval mode used when a search capability is unavailable.
Why we’re building it
Good context shouldn’t disappear with a chat.
We kept running into the same problem: increasingly capable models, but project context that had to be rebuilt for every conversation.
PRISM grew from wanting sources, decisions and disagreements to outlast the chat and remain useful as models change.
Open source is the intended direction
Built privately. Intended to become open.
We plan to release PRISM once its architecture and interfaces are stable enough for others to build on.
The project is still in development, with release timing to follow as the work matures.
Durable Project Sources underpin the derived Project Brain. Persistent Memory and a Context Builder are planned layers. Conversation is immediate context supplied by the connected agent. This is a research direction, not a claim that PRISM already implements a personal world model.
ConversationImmediate context from the agent
MemoryPersistent context and experience
Planned
Project BrainDerived · Searchable · Rebuildable
Project SourcesDurable · Versioned · Traceable
Context BuilderPlanned
PRISM is informing the foundations. AVeMe’s personal world model remains a future research direction.
A few useful distinctions
Questions about PRISM.
Where it runs, what it preserves and where the work is heading.
Can we use PRISM today?
We’re developing and testing PRISM at Comans. A public release has not been announced. Talk to us about your project and the work you would like to explore.
Does everything stay on my computer?
Source storage, the Project Brain and retrieval are designed to operate locally. If you connect a cloud model, the evidence included in its request can leave the local environment.
Is PRISM a model or a vector database?
PRISM works alongside AI models. It preserves and retrieves project evidence; the model reasons from it. Vector search is one retrieval method inside the wider architecture.
Why keep the original files?
Parsers, indexes and models change. Preserving the original source material means the derived Project Brain can be rebuilt without losing the evidence it came from.
How does this relate to AVeMe?
PRISM focuses on project knowledge. AVeMe explores longer-term personal context and temporal experience. Lessons from PRISM inform that research; they do not mean the future AVeMe system is already implemented.
Keep the conversation going
Your models will change. Your knowledge should last.
We welcome conversations with teams working on project knowledge, AI agents and evidence they can trace.
