Also in this release
MemHub + APIFeatures14
- Mistral OCR for document parsing, plus DOCX parsing with image descriptions.
- Document citations in team chat.
- Artifacts and documents take tags, which are searchable and patchable.
- Agent runs get a per-fire isolated sandbox on Fargate and Lambda.
- Ingest a document straight from a public URL.
- MCP server — brain tools: create, list, share, and brain-scoped recall.
- MCP server — workspace-aware tools, including
list_orgs and an org_id argument across the surface. - MCP server —
import_conversation brings a Claude Code conversation in wholesale. - MCP server —
delete_artifact, artifact tags, and lineage/diff queries. - MCP server —
search_memory now reaches the documents substrate. - APIFacts can expire.
- API
created_at time filtering and a since window on memory search. - APIArtifact versioning through
parent_id. - APIHybrid search — dense and lexical combined with reciprocal rank fusion.
Improvements1
- The unused personal-skills feature is removed end to end.
Fixes4
- Tags — tag columns use a Postgres
ARRAY so tag-filtered search works. - MCP server — RFC 9728 metadata is served at the full endpoint path.
- Brains — creation status surfaces to the frontend rather than looking stuck.
- APIEmbeddings — per-input tokens are capped at the embedding client.
Security1
- Document ingest — resource bounds against malicious uploads.