FINOS Labs adds konspekt, an open standard for portable AI decision records

FINOS has announced the addition of konspekt to its FINOS Labs program, an open-source contribution submitted individually by its maintainer, Denis Urusov. It is a proposed standard for creating a portable, human-readable record of the decisions, open questions, artifacts and provenance generated when teams work with generative AI tools over time and across different platforms.

The problem it seeks to solve is concrete and increasingly common in software engineering environments that use AI intensively: every working session with an AI assistant produces valuable decisions, reasoning, definitions of work and artifacts, but that knowledge ends up scattered across chat transcripts tied to the original platform. This creates two systemic problems that the article explicitly identifies. First, context fragmentation: reconstructing a project's context when switching tools, upgrading AI models or onboarding a new team member is costly and error-prone. Second, a governance gap: there is no independent, durable record documenting where a particular conclusion came from or who approved it. For regulated financial institutions, where provenance traceability and human accountability are strict requirements, this gap becomes a first-order governance problem.

konspekt's proposal consists of modeling a project's state as a compact, typed knowledge graph made up of a few node types —Concept, Noteworthy, Artifact and Waypoint— connected through a single edges table. Everything is stored in human-readable plain text files, which allows the graph to move cleanly between different AI platforms with a simple copy and paste. Each fact is stored only once, so inventories are queried dynamically instead of being maintained by hand.

The article highlights two properties that make konspekt a design especially suited to regulated financial institutions. The first is what it calls a 'human-curated review boundary': although an AI model can propose entries to the graph, only a person can accept them into the permanent record. Each entity retains verbatim provenance linking to the exact source text it came from, generating an explicit audit trail. The second property is that konspekt defines itself as an 'upstream' open format, that is, it sits above automated memory systems (the text cites Mem0 or Zep as examples) rather than competing with them: a human-curated konspekt instance offers a clean, validated record that those memory engines can reliably consume.

Among the technical features listed in the announcement are: a typed knowledge graph stored in flat files that can be reviewed and diffed as if it were source code; a 'propose and accept' workflow in which the model proposes entries but only a person accepts them, establishing an auditable review boundary; a small vocabulary of 'human authority verbs' —pin, validate, refute, resolve, abandon, lift— covering the override decisions an assistant should not make on its own; content-addressed provenance, so that each entry links to its verbatim source text through a verifiable content hash; optional 'persona' layers, role profiles that configure the operating instructions an assistant follows and the conformance checks applied to an instance, enabled per project with a simple flag (the 'engineer' role is presented as the pilot persona, logging the commands executed by the AI to maintain an auditable trail of the actions taken); and, finally, an open, vendor-neutral design, with no services to run and no proprietary runtimes, plus a dependency-free scaffolder that allows konspekt to be added to any existing repository.

According to the announcement, this approach connects directly with FINOS initiatives focused on AI governance, information provenance and vendor independence for financial sector infrastructure, an area in which FINOS —the open-source foundation focused on financial services— has been working through various projects and technical groups.

As for its current status, the project is in the 'FINOS Labs' phase, which corresponds to the early stages of incubation within the foundation, and its stated next milestone is to secure a second independent implementer. The code is available in the github.com/finos-labs/konspekt repository, where the specification lives in the spec/ folder and there is a working instance documenting konspekt's own development in .konspekt/instance/. To try it out, the project itself (which requires Node 18 or higher) can be added to a repository by running the setup/init.mjs script with the project's name and objective parameters. The reference documentation includes a README.md and a WHITEPAPER.md with the model and its rationale, along with the spec/data-model/ and spec/architecture/ folders for technical detail, and a CONTRIBUTING.md for those who want to take part. The community coordinates through a mailing list dedicated to konspekt and the FINOS Slack channel.

Taken together, the announcement fits into a broader trend seen across the sector: as generative AI becomes embedded in engineering workflows, proposals for open formats are emerging that seek to avoid dependence on a specific vendor and to ensure that decisions made with AI assistance are documented in a verifiable way, something especially sensitive in regulated sectors such as finance.