Memory as text for LLMs works well enough when the truth can be plainly stated and doesn't change, but I kept running into the same problem with LLM wikis, markdown, vector stores - I just didn't trust them, and I didn't trust my agent to properly state its evidence for a claim, its reason for changing something, or rebuild its semantic connections the same way between sessions from a text corpus. I found myself constantly asking my agents to re-verify what was already in my wikis, so what was really the point of the wiki in the first place?

I built Cruxible to fix that. It's a governed state layer that human/agent teams operate and iterate on together: you model what's true for your domain in an explicit, typed ontology, and the runtime enforces how state may change. It's designed to be unobtrusive: the state compounds as the exhaust of normal agent work, not as an ontology someone has to manually comb through.

Repo is here: https://github.com/cruxible-ai/cruxible

- All state that can be built deterministically is built deterministically. When judgment is needed, Cruxible includes strict governance review queues with full attribution receipts for all changes or proposals. Only proposals that are approved get minted into state.

- Every read is deterministic and executes outside the LLM. Agents can use a query engine to traverse the state knowledge graph or use configured named queries for repeating reads.

- The model is executable, you can read the state programmatically and do whatever logic you want with it. For example, in one of our bundled kits, the state models a company's IT structure deterministically using pinned source artifacts, and a workflow uses that structure to propose the assets that could be affected downstream. The query for this looks something like: KEV entry affects product → asset runs product → asset owned by team. The agent also gets the provenance for each edge and the original sources don't have to be referenced again, so no RAG once truth is minted into state.

- Reads are reproducible, and every piece of custom logic is pinned, allowing for exact snapshots and full reproducibility, no more guessing why something changed from a commit message on a md file.

I believe that hard state models of an organization's internal truth are going to be key internal assets and IP in the AI age, so all this is designed to be runnable locally through an isolated daemon.

The README demo seeds a supply-chain world and walks the loop in ~3 minutes so you can see it in action, or you can ask your agent to build you your own kit using the bundled skills. Apache-2.0, Python, SQLite under the hood (a single state.db per instance, under the daemon's state directory), with a read-only inspection UI in a separate repo (link in the README).

I'd really value feedback of any kind, even if it's just how you and/or your team are maintaining consistency about what's true. Contributions welcome!

Show HN: Cruxible – Open-source governed truth layer for AI agents