Interoperability stance
Agent memory is a crowded, fast-moving market — and Kontour is not in it. We build no memory platform. That's a recorded do-not-build guard, not a roadmap gap: sidecar state and trust bundles are evidence about work, not memory about the world, and they operate alongside whatever memory stack your agents already run.
Memory tells the agent what it knows. Kontour proves what its answers stood on.
The two layers
A memory layer answers "what does the agent know going in?" Kontour answers "what did the answer stand on coming out?" The first is a context concern; the second is a trust concern. Standardizing the first — as knowledge formats now do — makes the second question more acute, not less: once everyone can assemble context, "but is the answer grounded?" is the question left standing.
e.g. OKF (Open Knowledge Format)
Standardizes the knowledge an agent consumes — context going in. It is deliberately silent on trust: no integrity, provenance, or freshness semantics. That's the seam, not a conflict: a knowledge bundle is a clean source for grounding, and a trust layer is built to supply what any such format leaves out.
e.g. Context Lattice
Owns what the agent remembers across sessions — retrieval, salience, forgetting. Kontour never sees or stores that memory. It records what the work stood on when the agent acted, whatever memory system fed it.
harness hooks writing to your own store
Rolls memory by hand with runtime hooks. Kontour's hooks ride the same harness surfaces without claiming them — sidecar state is workflow evidence under .kontourai/, not a memory retrieval index. What the hooks inject is workflow steering, never remembered knowledge.
Honest scope: these are examples of the architectural pattern, not tested integrations or partnership claims. We describe how the layers compose; we don't certify vendors.
How they compose
In practice the seam is simple: your memory layer decides what context the agent sees; the harness does the work; Kontour's hooks capture what actually ran and what the claims stood on — real exit codes, tamper-evident records, recomputabletrust.bundle artifacts. Swap the memory layer and nothing on the trust side changes, because nothing on the trust side ever depended on it.