v0.4.2 — docs alignment and Gemini CLI prompt pickup

Highlights

• Gemini CLI now reads AGENTS.md natively. Both the bundled examples/clients/gemini.settings.json fixture and the in-repo .gemini/settings.json ship "context": { "fileName": ["AGENTS.md", "GEMINI.md"] }, so Gemini CLI picks up the same project prompts that Codex, Cursor, and Claude Code already use — without forcing a duplicated GEMINI.md mirror.
• Install URLs realigned in README.{md,uk.md,ru.md}, examples/clients/SMOKE_CHECKLIST.{md,uk.md,ru.md}, and examples/clients/codex.config.toml so users following the docs install the actual current build instead of an older v0.3.1 snapshot.
• Upgrading section added to all three READMEs covering uv tool install --reinstall and a one-time ~/.gemini/settings.json migration for existing users.
• SMOKE checklists gained a Prompt pickup step in the Gemini CLI client checks that warns operators to preserve the context block when merging the fixture into an existing settings.json.
• Lint false-positive TODO filed (.planning/todos/2026-04-28-lint-false-positives.md) for the next code release: ASCII-only title normalization that collapses non-ASCII H1s into untitled, and broken_link reports for files inside DEFAULT_IGNORED_DIR_NAMES (e.g. benchmarks/latest.md) that physically exist on disk.

Upgrade

uv tool install --reinstall git+https://github.com/Lexus2016/turbo_quant_memory@v0.4.2

If you already have ~/.gemini/settings.json from before v0.4.2, also merge this block once so Gemini CLI starts reading AGENTS.md alongside GEMINI.md:

{
  "context": { "fileName": ["AGENTS.md", "GEMINI.md"] }
}

Notes

• Code paths are unchanged; this is a docs/integration patch.
• All 93 tests pass against the published source.

v0.4.1: Auto-failover when primary dies

Critical fix

When the first turbo-memory-mcp process (the primary holding the sentence-transformers model and LanceDB handles) shut down, the remaining proxy processes in other MCP clients kept stale RPC connections and tqmemory became silently unavailable for those clients. v0.4.1 restores cross-process memory sharing to work transparently across primary restarts.

What's new

• Automatic proxy failover. Proxies detect a lost primary via PrimaryUnreachable on the next RPC call and transparently re-bootstrap: one surviving proxy atomically claims the lockfile and promotes itself (starting its own DaemonListener), while every other orphaned proxy reconnects to the new primary. No MCP client restart required.
• Phase-aware RPC error handling. Connect/send-phase failures are translated into PrimaryUnreachable (safe to replay), while mid-call failures (send succeeded, recv failed) surface to the host unchanged so non-idempotent tools like remember_note are never silently duplicated.
• Serialized failover. Concurrent PrimaryUnreachable detections in the same process are coalesced under the runtime's state lock, preventing the second thread from racing into the bootstrap before the first thread's listener starts accepting — which previously could have let two threads both install primary state inside one process.

Upgrade

uv tool install --force --reinstall 'git+https://github.com/Lexus2016/turbo_quant_memory@v0.4.1'

No on-disk state changes; JSON notes, LanceDB tables, and Markdown manifests from v0.4.0 are fully backward-compatible.

v0.4.0 - Singleton daemon for cross-process memory sharing

Transport-level release that coordinates multiple MCP clients on the same machine through a single primary process. No migration required — all on-disk formats (JSON notes, LanceDB tables, Markdown blocks, manifests) are unchanged.

Highlights:

• Singleton daemon: only one turbo-memory-mcp process per machine holds the sentence-transformers model and LanceDB handles. Additional MCP-client launches become thin stdio↔socket proxies that forward tool calls to the primary.
• Cross-platform coordination: Unix/macOS uses an AF_UNIX socket under the system temp dir (short path, 0600 perms); Windows uses a named pipe scoped to the current user. Authenticated via a 32-byte random authkey stored in ~/.turbo-quant-memory/.daemon.lock.
• Lazy imports in retrieval_index so proxy processes do not pay the ~470 MB import cost of PyTorch / LanceDB / PyArrow when they only forward RPC. Measured savings: ~1 GB RSS for four concurrent MCP clients (primary ~530 MB, proxies ~50 MB each vs ~437 MB each before).
• Strict no-retry RPC semantics: transport failures mid-call surface as ConnectionError so non-idempotent tools (remember_note, promote_note, deprecate_note, index_paths) can never be silently duplicated.
• Escape hatch: TQMEMORY_DAEMON_DISABLE=1 falls back to per-process mode (previous v0.3.x behaviour).

Upgrade notes:

• Run uv tool upgrade turbo-memory-mcp (or reinstall via pip).
• Restart every MCP client so both ends of the stdio bridge pick up the new coordination protocol.
• Existing project memory is preserved — the primary reads the same ~/.turbo-quant-memory/ tree you had before.

v0.3.2 - .tqmemoryignore support and telemetry fixes

Patch release for indexing control, retrieval quality, and telemetry accuracy.

Highlights:

• add .tqmemoryignore support so index_paths can exclude duplicate workspaces, reports, and generated Markdown via glob patterns
• fix retrieval key point extraction so compact results split and preserve useful segments more reliably
• improve usage telemetry migration and byte/token savings accounting for compact and hydrated context
• document the new ignore workflow in English, Ukrainian, and Russian and cover it with ingestion and namespace tests

v0.3.1 - Shared-memory docs and Gemini CLI rollout

Patch release for cross-client continuity documentation and Gemini CLI support.\n\nHighlights:\n- document shared local project memory handoffs across Codex and Gemini CLI\n- add Gemini CLI client integration guidance, smoke coverage, and ready fixture\n- update live client tiers so server_info/self_test include Gemini CLI\n- refresh release-facing install/version references to v0.3.1

v0.3.0

Highlights

• Added version-aware markdown and retrieval manifests with automatic rebuilds when derived indexes become stale after an upgrade.
• Added persistent usage telemetry outside project/global memory, including estimated byte/token savings, milestone headlines, and optional USD estimates when TQMEMORY_INPUT_COST_PER_1M_TOKENS_USD is configured.
• Hardened retrieval behavior with project-safe default scope, stale freshness detection, removed-root pruning, incremental sync with full-sync fallback, Unicode-aware lexical fallback, and cached server_info() storage snapshots.
• Refreshed README, smoke coverage, and client rollout examples for v0.3.0.

Verification

• uv run ruff check src tests scripts
• uv run pytest -q
• uv run python scripts/smoke_test.py

v0.2.4 - Knowledge-base linting and release refresh

Highlights

• Added new MCP tool: lint_knowledge_base(...) for markdown knowledge-base hygiene checks.
• Added deterministic lint diagnostics for broken internal links, orphan candidates, duplicate normalized titles, and Obsidian-style wikilinks.
• Extended server/tool contracts and self-test catalog to include the new lint tool.
• Added dedicated lint test coverage and updated smoke flow to validate lint behavior end-to-end.

Documentation

• Updated README files in English, Ukrainian, and Russian.
• Updated technical specs and client smoke checklists for the new tool and updated install commands.

Versioning

• Bumped package version from 0.2.3 to 0.2.4.

v0.2.3

Turbo Quant Memory v0.2.3

• splits user-facing repository docs into clean English, Ukrainian, and Russian files instead of mixed bilingual pages
• adds localized versions of Memory Strategy, Technical Specification, Client Integrations, benchmark report, and client smoke checklist
• refreshes benchmark artifacts against the new 17-file documentation corpus and keeps the readable summary SVG layout in the generator
• bumps the package and install contract to v0.2.3 and updates tests, examples, and release-facing commands accordingly

v0.2.2

Turbo Quant Memory v0.2.2

• fixes package metadata so the published package version now matches the Git tag, install contract, and server version
• keeps the v0.2.1 memory hygiene, agent instruction, benchmark, and SVG fixes intact
• is the recommended release for installation and agent rollout

v0.2.1

Turbo Quant Memory v0.2.1

• improves default memory hygiene by skipping low-signal folders such as .planning, .serena, and generated benchmark reports during project-root indexing
• adds explicit Codex and Claude Code memory workflow instructions through AGENTS.md and CLAUDE.md
• cleans README benchmark snapshots and localized install docs
• fixes overflow in localized README hero SVG assets
• refreshes benchmark artifacts and keeps the release install contract pinned to v0.2.1