Releases: steelbrain/agent-journal
Releases · steelbrain/agent-journal
Release list
v0.2.0 - Now with --dump!
Added
agent-journal --dumpwrites the current project's memory as markdown: one
file per entity (statements grouped under it, invalid ones marked) in
.agents/knowledge-base/and one file per UTC day in.agents/journal/.
Every file carries an autogenerated do-not-edit header. Each run rewrites
the.mdfiles in those directories so the dump stays in sync with the
database; other files are left untouched.
Changed
- Schema v2: search indexes are project-scoped. The vector index is
partitioned by project with owner type and status metadata, and the FTS
tables carry project/status columns, so search recall budgets apply per
project and per store instead of to one global pool, and invalidated
tombstones stop consuming recall slots unlessinclude_invalidis set.
Existing databases migrate automatically; vectors are copied without
re-embedding. The migration takes the write lock up front and re-checks the
schema version under it, so concurrent first starts after an upgrade cannot
race, and the rebuild is idempotent in any case. kb.upsert_entitytitle changes now re-embed the entity's statements so
semantic search reflects the new title (previously only keyword search was
re-synced).memory.recentpages via SQL ordering and aggregates instead of loading
whole tables into memory.- Deletion audit journal entries are now keyword-searchable.
Fixed
kb.deleteon an entity no longer fails with a foreign-key error: it now
cascades to the entity's statements and relationships, cleaning their search
index entries, journal links, and inbound redirects, records the cascade in
the audit journal entry, and reportscascaded_statementsand
cascaded_relationshipsin the result.memory.recentno longer silently drops records that share acreated_at
with a page boundary. Paging now uses a compound cursor: pass the returned
next_beforeandnext_before_idback asbeforeandbefore_id.- A busy checkpoint or vacuum after
kb.deleteis no longer reported as a
failed deletion; the result now carriesvacuum_completedinstead.
v0.1.0 - Initial Release
Added
- Initial release of the
agent-journalMCP server: a local, project-scoped
knowledge base and journal for coding agents. - Knowledge base of immutable, confidence-tracked statements attached to
entities, with immutable edits and invalidation instead of in-place updates. - Append-only journal recording what was done and what was proven, with links
to knowledge-base records. - Search over both stores via SQLite FTS5 and sqlite-vec, backed by local
embeddings, with humanized relative timestamps on results. - MCP tools:
memory.search,memory.get,memory.recent,memory.stats,
memory.guide,memory.agents_md_snippet,kb.upsert_entity,
kb.add_statement,kb.edit_statement,kb.invalidate,kb.delete, and
journal.append. - Automatic project resolution from git remote, worktree, or working directory,
with per-repo config files under the XDG app config directory. --mcpflag to run as a server; bare invocation prints help and--version
reports the version.