Skip to main content
The harness gained a more practical memory loop: recall now spans code and project memory, longer runs can pack older context, and generated project instructions can be checked for drift before they are rewritten.

​
Semantic code recall

voss recall now searches code and project memory together. Code hits include source labels and line information, while memory hits stay labeled as memory. The same recall path is also available to the harness through the code_recall tool. The code index is rebuildable cache state. If vector search is unavailable, Voss can still use lexical recall instead of failing the workflow.

​
Budget-aware context

Long-running harness work now has conservative context packing. Recent work stays detailed, older replay can be reduced to structural summaries, and Voss writes estimated original vs. packed token counts to the session’s token-savings.jsonl. The escape hatch is explicit: 
voss do "inspect the auth flow" --no-pack
Savings are estimates, not billing guarantees.

​
Managed sync checks

voss sync --check gives generated workflow docs and managed instruction sections a read-only drift gate. It is intended for CI and pre-commit style checks where plain voss sync would be too mutating.

​
SDK proof suite

The internal proof suite is exercising the Python, TypeScript, Go, and Rust consumer paths against voss serve. This is proof infrastructure, not a public SDK launch claim yet; the remaining live checkpoint still determines when those flows should be promoted in user-facing docs.

​
Next memory work

Global memory is planned as an explicit, local, curated store for facts that should cross projects. External markdown/docs ingest remains planned after that and stays read-only by design.