onfiala/dots-brain
Install and operate self-hosted MCP memory for explicitly saved notes.
Changelog
Unreleased
- Extend the existing synthetic retrieval runner with separate development and held-out fixtures, three deterministic seeds, current-revision/source/passage guards and explicit no-answer metrics. Keep held-out execution outside automatic CI and preserve the original small baseline.
- Add optional bounded Linux capacity measurement with 450 synthetic records, approximately 30,000 chunks, three query batches and an enforced 2 GiB test-unit budget. These are measurement tools; retrieval behavior remains unchanged.
1.0.2
-
Read canonical memory totals, source groups and semantic index counts from one SQLite snapshot in
memory_status, including project-scoped requests. Concurrent saves, indexing, revisions and deletions no longer mix counts from different moments. -
Preserve read-only status behavior without blocking concurrent WAL writers; retain standalone component status calls and semantic-disabled operation.
-
Retain schema v2, dependency versions, model artifacts, search behavior and existing credentials. No data migration or index rebuild is required.
-
Run the pinned real embedding model in a dedicated CI job with verified model caching and a receipt that rejects skipped tests. Publish a frozen synthetic CZ/EN full-text/hybrid relevance baseline with source and environment hashes. Retrieval behavior, dependencies and schema remain unchanged.
-
Accept technical URL usernames, explicit credential placeholders and lowercase
sk-learnreferences without weakening recognized credential rejection. Return safe detection categories alongside field labels on rejected writes. -
Preserve unrelated URL query pairs and fragments during minimization, and make supported redacted values stable on recheck. Add synthetic Store, audit and MCP regressions. Capture and CORTEX modules remain disabled; no schema or dependency changes are required.
1.0.1 — 2026-11-08
- Isolate managed Python module launches from caller-controlled import paths and safely reconnect exactly owned bridges across installation environments.
- Require audit receipts to reference an existing actor/project intent and allow only one distinct receipt, while preserving review progress across historical anomalies and adapting oversized audit pages to the review budget.
- Reject malformed credential identities with safe JSON errors and keep status counts consistent during concurrent writes and indexing.
- Strengthen the responsiveness regression to require an acquired SQLite lock and a waiting writer, with cleanup on failed assertions.
- Retain schema v2, existing dependencies, scoped credentials, OAuth history and model artifacts. Capture and CORTEX modules remain disabled.
1.0.0 — 2026-11-08
- Establish a stable release for the existing Linux memory core: explicit notes, revisions, project-scoped access, full-text search and optional local embeddings.
- Document pinned release installation, checksum verification, clean installed-wheel acceptance and compatible binary rollback. Schema v2 and runtime behavior remain unchanged from 0.4.4.
- Clarify the fixed 30-day OAuth grant lifetime and owner-approved re-pairing; token refresh does not extend authorization or remove stored memories.
- Update the locked certifi dependency to 2026.10.19, matching the maintained deployment. No paid API integration or hosted store is added.
- Capture and CORTEX modules remain disabled and excluded from core artifacts. Full host-loss recovery and automatic assistant memory use remain unsupported.
0.4.4
- Make the default systemd example work with core dependencies. Document the explicit semantic-extra and model preparation path for its service owner.
- Verify fresh core-only wheel installations in the Linux Python 3.11/3.12 matrix, including real stdio/HTTP MCP, template startup, migration and restore. Reject source imports, mismatched versions and excluded module packages.
- Require current locked metadata, main ancestry and installed-wheel checks in release preparation before uploading artifacts to an existing draft.
- Reject incomplete indexing or fulltext fallback in the semantic stress harness, including when fallback still returns the expected synthetic records.
- Preserve an existing empty or unknown source disable barrier during restore activation; only a valid retry of the same recovery may proceed.
- Validate package, runtime, plugin, lockfile, README and changelog versions together. Schema v2 and dependencies are unchanged.
0.4.3
-
Give replacement FTS documents a fresh row ID before deleting the old entry, preserving secure deletion after updates to long records.
-
Exercise large request buffering through nginx in Linux CI and document safe gateway validation and temporary-directory ownership checks.
0.4.2
- Replace obsolete verification history with reproducible source, installed-wheel, Linux, retrieval, and deployment checks.
- Correct release workflow and bootstrap resume guidance; consolidate historical deployment pointers into current operating documentation.
- Align package descriptions, contributor commands, and troubleshooting with the implemented behavior.
- Report history import, continuous and provider-wide capture as disabled in memory status. Storage, retrieval, schema, and dependencies are unchanged.
0.4.1
- Preserve a concurrently replaced credential or export file when creation fails.
- Report incomplete output cleanup while retaining the original error and keeping private diagnostic details out of CLI output.
- Keep semantic chunks in deterministic memory order so candidate pruning retains the best body passage for a title match.
- Correct setup guidance: capture and the CORTEX connector remain disabled and excluded from the core package. Schema v2 and dependencies are unchanged.
0.4.0
- Retain schema v2 and the explicit released v1 migration. Add read-only version checks, FTS secure deletion, snapshot/recovery validation and writer/HTTP leases.
- Preserve OAuth client/grant provenance after invalidation; bind public policy to its listener and make bulk access changes explicit offline maintenance.
- Reject known secrets, warn on uncertain content, retain successful writes when later auditing fails, and report unknown remote outcomes without automatic retry.
- Return the source body/revision with title matches; make indexing counts and FTS fallback explicit and isolate bridge sessions for concurrent calls.
- Publish managed state after readiness; bind process control to its data path. Add read-only connect with failure rollback and explicit supervisor endpoints, independent doctor checks, timestamps and one startup log rotation.
- Add ownership-aware disconnect and uninstall that preserve unrelated client settings, changed credentials and retained memory data.
- Add schema-v2 audit review with checkpoint continuity and operator diagnostics.
- Exclude disabled CORTEX and capture modules and their tests from the core distribution while preserving their source-only characterizations.
- Add installed-wheel verification using a released synthetic v1 fixture and real stdio MCP.
This version uses selected changes from the later experimental branch; its v3/v4 schemas and additional command surfaces are not part of this v2 result.
0.4.0-alpha.1 — 2026-10-09
The existing tag is retained; it is not the batch-1 candidate.
Added
- Add project-scoped identities, authenticated revision authorship, passage embeddings, a sanitized action audit, and an experimental JSONL snapshot importer.
- Add optional CORTEX context and publication of a selected memory revision.
- Add validated backups, staged restores, deletion reconciliation, and the operator audit review helper.
Breaking
- Schema v2 requires an explicit offline migration from v1.
memory_forgetrequires the caller's observedexpected_revision.
0.3.0-alpha.2 — 2026-10-04
- Tighten OAuth issuer validation and supersede the alpha.1 release candidate.
0.3.0-alpha.1 — 2026-10-04
- Add optional OAuth discovery, dynamic registration, PKCE, scoped owner approval, refresh, and revocation using the MCP SDK.
0.2.0-alpha.2 — 2026-10-02
- Add previewable uninstall and disconnect while retaining memories and unrelated client settings. Require explicit resume after uninstall.
0.2.0-alpha.1 — 2026-10-02
- Add managed Linux startup, local client adapters, bridge verification, and bounded storage, HTTP, and semantic stress scripts.
0.1.0-alpha.2 — 2026-10-02
- Bound and retry first-run WAL lock contention without losing existing data.
- Let both Python CI jobs finish when one matrix job fails.
0.1.0-alpha.1 — 2026-10-01
- Introduce SQLite source records, revisions, project-filtered full-text search, exports and deletion suppression.
- Add six MCP tools, stdio and authenticated loopback HTTP, local credentials, a bridge and machine-readable setup commands.
- Add pinned multilingual CPU embeddings, background indexing and hybrid search.
- Add the bootstrap and onboarding skill, release validation and Python 3.11/3.12 CI.
Known limits
- Public ingress, provider account setup, continuous conversation capture, native client UI activation, off-host recovery, and reboot acceptance are not verified.