Skip to content

Releases: steelbrain/agent-journal

v0.2.0 - Now with --dump!

Choose a tag to compare

@steelbrain steelbrain released this 04 Jul 23:43

Added

  • agent-journal --dump writes 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 .md files 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 unless include_invalid is 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_entity title changes now re-embed the entity's statements so
    semantic search reflects the new title (previously only keyword search was
    re-synced).
  • memory.recent pages via SQL ordering and aggregates instead of loading
    whole tables into memory.
  • Deletion audit journal entries are now keyword-searchable.

Fixed

  • kb.delete on 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 reports cascaded_statements and
    cascaded_relationships in the result.
  • memory.recent no longer silently drops records that share a created_at
    with a page boundary. Paging now uses a compound cursor: pass the returned
    next_before and next_before_id back as before and before_id.
  • A busy checkpoint or vacuum after kb.delete is no longer reported as a
    failed deletion; the result now carries vacuum_completed instead.

v0.1.0 - Initial Release

Choose a tag to compare

@steelbrain steelbrain released this 24 Jun 23:53

Added

  • Initial release of the agent-journal MCP 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.
  • --mcp flag to run as a server; bare invocation prints help and --version
    reports the version.