The Challenge: Scattered AI History and Unauditable Decisions

As teams increasingly do substantive work through generative-AI conversations, every session produces something worth keeping: decisions, reasoning, working definitions, and the artifacts themselves. Today, that state is scattered across chat transcripts and tied to its original platform. This creates two major systemic issues for financial institutions:

Context Fragmentation: Knowledge does not cleanly move across tools. Reconstructing project context when switching tools, upgrading models, or onboarding team members is costly and error-prone.

Governance Gaps: There is no independent, durable record of where a given conclusion came from or who approved it. For regulated financial institutions, where provenance and human accountability are strict requirements, that gap is a major governance problem.

The Solution: A Portable, Human-Curated Record of Project State

konspekt defines project state as a compact, typed knowledge graph—using a few core node types (Concept, Noteworthy, Artifact, Waypoint) connected by a single edge table. Stored entirely as plain, human-readable files, the graph moves cleanly between AI platforms via simple copy-paste, storing every fact once so inventories are dynamically queried rather than manually maintained. Two key properties make konspekt tailored for regulated financial institutions:

Human-Curated Review Boundary: While an AI model may propose graph entries, only a human can accept them into the permanent record. Every entity carries verbatim provenance linking back to its exact source text, creating an explicit audit trail.

Upstream Open Format: As a zero-dependency open format, konspekt sits upstream of machine memory systems (such as Mem0 or Zep) rather than competing with them. A curated konspekt instance provides a clean, human-validated record that downstream memory engines can reliably consume.

This approach aligns directly with FINOS initiatives focused on AI governance, provenance, and vendor independence for financial infrastructure.

Key Features

• Typed knowledge graph in plain files: A small schema of typed nodes and a single edge table, stored as human-readable text that diffs and reviews like source code.
• Human-in-the-loop propose→accept: A model may propose entries; only a person accepts them into the record, creating a clear, auditable review boundary.
• Human authority verbs: A small vocabulary a person issues over the graph (pin, validate, refute, resolve, abandon, lift), covering the override moments an assistant must not decide on its own.
• Content-addressed provenance: Every entry links to its verbatim source text, addressed by content hash, so any statement can be checked against exactly what it came from.
• Persona layers: Opt-in role profiles that shape the operating instructions an assistant follows and the conformance checks applied to an instance, switched on per project with a single flag. The engineer role is the pilot persona; it records the commands the AI executes, keeping an auditable trail of the actions taken during a project.
• Open and vendor-neutral: Portability is achieved by keeping project state as plain, human-readable files that move between generative-AI tools by copy-paste, with no service to run and no proprietary runtime. A zero-dependency scaffolder drops konspekt into any existing repository.

Get Involved

• Explore the Code: github.com/finos-labs/konspekt. The standard lives in spec/, with a working instance that tracks konspekt's own development in .konspekt/instance/.
• Try It: From a project root (Node 18+), run node setup/init.mjs --name "My Project" --goal "what the project is for" to scaffold konspekt into your repo; see setup/README.md.
• Read the Standard: README.md and WHITEPAPER.md for the model and rationale; spec/data-model/ and spec/architecture/ for the details.
• Contribute: CONTRIBUTING.md for the process. The project is at the FINOS Labs stage, and its stated next milestone is a second, independent implementer, a concrete place to start.
• Connect with the Team: Join the konspekt mailing list and the FINOS Slack at finos-lf.slack.com to reach the maintainer and community.