Added

• Per-folder memory is now first-class: the deployed agent skill explains that runtime and data root (MEMORIA_HOME) are separate, so any clean folder can get its own memory by running setup there. Deployed SKILL.md adds self-location of the data root, a fail-closed "check before write" flow, and a new-folder setup walkthrough.
• resolveMemoriaHomeInfo() reports how MEMORIA_HOME was resolved (env / detected / fallback). getMemoriaHome() delegates to it with unchanged behavior and return type.

Changed

• doctor no longer reports MEMORIA_HOME as always healthy. When the home was resolved by silent fallback (folder never set up, no env), the check now fails with an actionable fix hint, and doctor --json exposes a homeSource field.
• The deployed agent skill now installs to <memoria-home>/.agents/skills/memoria/ (was <memoria-home>/.agents/memoria-memory-sync/) and is named memoria, so agents discover it as the memoria skill in active_skills. The repo-side source directory skills/memoria-memory-sync/ is unchanged.
• README / README.zh-TW / docs/INSTALL.md updated for the new deployed path, skill name, and per-folder memory model.

Fixed

• npm-installed setup now deploys the full agent skill. The files whitelist only published skills/memoria-memory-sync/deployed/, so the npm tarball was missing SKILL.md, scripts/, and resources/ — getBundledSkillSourcePath() found no SKILL.md and deployAgentSkill was silently skipped, leaving .agents/memoria-memory-sync/ undeployed. Whitelist now ships the whole skills/memoria-memory-sync/ directory.
• setup now logs a ✗ [skill] step when the bundled skill source is missing, instead of skipping the step with no output.

Added

• Claude Code adapter: wire Memoria into Anthropic's Claude Code via its hook system without writing any code. One CLI command (memoria adapter claude-code) handles both UserPromptSubmit (injects recall as additionalContext) and Stop (writes the just-completed turn from the transcript). Both fail-open so a Memoria outage never disrupts the agent loop.
• scripts/test-claude-code-adapter.sh end-to-end test, wired into ci.yml.

Changed

• HTTP server hot path (recall, recallTelemetry, stats) now reuses a cached SQLite connection across requests instead of opening + closing per call. Server SIGINT/SIGTERM handlers drain the pool via closeAllConnections().
• initDatabase() no longer inlines ALTER TABLE patches. Schema upgrades are tracked in a new schema_migrations table, with three numbered migrations corresponding to the previous inline patches.
• dist/ is now gitignored. Build artifacts (dist/cli.mjs, dist/install/memoria) are regenerated by pnpm run build and pnpm run release:package; prepublishOnly ensures npm publish still works. Removes ~21 k lines of binary-shaped diff from every commit.

Added

• npm publish target: @raybird.chen/memoria (scoped public package). npx @raybird.chen/memoria setup or npm install -g @raybird.chen/memoria now works on Linux / macOS / Windows.
• scripts/build.mjs build entry point that emits dist/cli.mjs with #!/usr/bin/env node shebang and executable permission, so npm-installed users get a working memoria binary.
• scripts/bump-version.mjs single-command version bump across package.json, src/cli.ts, install.sh, deployed skill, and docs/INSTALL.md.
• .github/workflows/release.yml — tag-driven release: pushing v* runs the pre-release checks, creates the GitHub Release with extracted CHANGELOG notes, and publishes to npm with provenance.

Changed

• package.json flipped from private: true to public (@raybird.chen/memoria) with a files whitelist (9-file, ~148 kB tarball).
• bin now points to dist/cli.mjs directly instead of the bash wrapper ./cli, so the npm install entrypoint works cross-platform.
• README / docs/INSTALL.md promote npm as Method A; no-clone tarball and repo dev mode move to Method B / Method C.
• RELEASE.md simplified to the new tag-driven SOP (release:bump → CHANGELOG edit → tag push → CI publishes).

Highlights

Domain module split — no breaking changes to CLI, HTTP API, or SDK.

Changed

• Split src/core/db.ts (2409 lines, 8 domains) into 11 focused modules under src/core/db/ (schema, session, source, wiki, lint, sync, telemetry, verify, prune-export, recall, mappers). All 32 public exports remain unchanged via src/core/db/index.ts.
• Split src/cli.ts (~890 lines, 16 commands) into a thin 50-line Commander registration shell, with each command extracted into src/cli/commands/ and shared helpers into src/cli/shared.ts, runtime.ts, preflight.ts.
• Updated AGENTS.md, README.md, and README.zh-TW.md to reflect the new directory structure.

Verification

pnpm run check
pnpm run build && node dist/cli.mjs --help
bash scripts/test-smoke.sh
bash scripts/test-bootstrap.sh
bash scripts/test-no-clone-install.sh

Changelog

All notable changes to this project will be documented in this file.

[Unreleased]

[1.8.0] - 2026年04月11日

Added

• Deployed skill packaging for no-clone installs, including runtime-safe SKILL.md and REFERENCE.md deployment into <memoria-home>/.agents/memoria-memory-sync.
• Release-time validation for deployed skill version alignment, required asset completeness, and repo-only instruction leakage.

Changed

• setup now ships a local bin/memoria wrapper with the deployed skill so installed agents can execute skill workflows without a cloned repo.
• README, install guide, and release SOP now document deployed skill discovery, packaging guards, and no-clone release expectations.

[1.7.0] - 2026年04月07日

Added

• Compiled wiki workflows for Memoria, including raw source import, generated knowledge/index.md / log.md / overview.md, query file-back, and durable wiki lint findings.
• Focused wiki coverage in CI for ingest, build, query filing, and governance lint flows.

Changed

• Agent guidance, operations docs, release SOP, and skill instructions now treat the compiled wiki as a first-class runtime workflow.

[1.6.0] - 2026年04月01日

Added

• No-clone release packaging via pnpm run release:package, including a Linux x64 runtime tarball with packaged better-sqlite3 dependencies.
• Artifact-based installer flow and no-clone end-to-end coverage for preflight, setup --serve --json, remember, and recall.

Changed

• preflight and setup now distinguish repo mode from installed mode, so packaged runtimes no longer require pnpm or repo-local dependency installation.
• README, install guide, release SOP, and CI now document and validate the release-asset installation path.

[1.5.1] - 2026年04月01日

Added

• Native ESM adapter runtime regression test at scripts/test-adapter-runtime.sh and CI coverage for bootstrap + adapter runtime verification.

Fixed

• ./cli setup now resolves the project install directory from the CLI entrypoint instead of MEMORIA_HOME, so bootstrap setup no longer runs pnpm install in the wrong location.
• scripts/test-bootstrap.sh now resolves repo root correctly.
• BaseAdapter now constructs MemoriaClient without CommonJS require(), so URL-string configuration works under native ESM runtime.

Changed

• Release and operations docs now include the bootstrap + adapter runtime verification steps in the patch release SOP.

[1.5.0] - 2026年03月17日

Added

• Adaptive retrieval gate for trivial recall queries (greetings, emoji-only messages, short confirmations) when no explicit recall mode or memory-intent phrase is present.
• Import guardrails that suppress exact duplicate events within a session and derive a better session summary from higher-signal events when the provided summary is trivial.
• Lightweight scope isolation: sessions can carry optional scope, which defaults to project:<project> or global, and recall/index flows can filter by scope.
• Governance review command (memoria govern review) for deterministic surfacing of repeated decisions and skills worth extracting.

Changed

• Recall telemetry and stats now include route_mode=skipped when adaptive retrieval intentionally bypasses lookup.

[1.4.0] - 2026年03月03日

Added

• Time-decay scoring for recall: memory relevance now decreases with age using 1 / (1 + age/halfLife) (halfLife=90 days). Newer memories rank higher when token match is equal.
• Keyword recall (recallKeyword) now computes relevance scores and sorts by score instead of timestamp-only ordering.
• Recall hit tracking: recallTree updates last_synced_at on matched memory_nodes for stale detection.
• prune --consolidate-days <N>: merges old session nodes under the same topic node (keeps newest, removes ≥3 old children).
• prune --stale-days <N>: removes memory_nodes (level=2) never recalled and orphan sessions older than N days.
• prune --all now includes --consolidate-days 90 and --stale-days 180 by default.

Changed

• scoreNode() accepts optional timestamp parameter for time-decay weighting.
• recallKeyword() return type now includes score field.
• recall() in MemoriaCore no longer synthesises position-based scores; uses actual computed scores.

[1.3.0] - 2026年02月26日

Added

• Tree memory index schema (memory_nodes, memory_node_sources) and memoria index build command.
• Tree/hybrid recall mode with explainable reasoning_path metadata.
• Recall routing telemetry with aggregated stats and raw API endpoint (GET /v1/telemetry/recall).
• MCP sync cursor state (memory_sync_state) and post-ingest cursor commit script.
• Incremental MCP payload mode (MEMORIA_MCP_PAYLOAD_MODE=incremental, default) with compatibility fallback to full.

Changed

• memoria sync now auto-builds incremental tree index by default (MEMORIA_INDEX_AUTOBUILD=0 to disable).
• Hybrid MCP flow now supports true no-op second sync (entities/relations unchanged when no deltas).
• memoria stats now reports 7-day recall routing quality (fallback rate, route counts, latency, hit count).
• Operations and MCP docs updated for tree recall observability and incremental sync controls.

[1.2.0] - 2026年02月14日

Added

• memoria sync --dry-run for validation and write-preview without mutating files.
• memoria stats command for sessions/events/skills summary and top skills.
• memoria verify command for runtime/schema/writeability validation with --json output.
• Agent Skill at skills/memoria-memory-sync/SKILL.md with references, resources, and helper scripts.
• Hybrid MCP bridge automation (run-sync-with-enhancement.sh) for optional mcp-memory-libsql ingestion.
• Installer preflight checks for common container tools (node, pnpm, npm, git, unzip, python3).
• Explicit path env support: MEMORIA_DB_PATH, MEMORIA_SESSIONS_PATH, MEMORIA_CONFIG_PATH.
• Bundled CLI build output at dist/cli.mjs with pnpm run build.

Changed

• Session JSON parsing now validates schema with clearer error messages.
• README now documents dry-run and stats command usage.
• Documentation now includes agentskills integration and MCP/libSQL auto-ingest workflow.
• Installer now supports npm fallback and --minimal/--no-git modes.
• CLI launcher now works with pnpm or npm (no hard pnpm requirement at runtime).
• ./cli now prefers built artifact (dist/cli.mjs) when available.

[1.1.1] - 2026年02月13日

Added

• GitHub Actions CI workflow for TypeScript and shell validation.
• Smoke test script at scripts/test-smoke.sh for end-to-end init + sync verification.
• SECURITY.md with private reporting and open-source data safety guidance.
• MIT LICENSE file.

Changed

• Documentation now includes open-source safety guidance and sample sync test flow.

[1.1.0] - 2026年02月13日

Added

• TypeScript CLI (cli, src/cli.ts) with init, sync, and doctor commands.
• Sample session file at examples/session.sample.json.

Changed

• Install and hook flow updated to TS-only runtime.
• Path handling aligned around MEMORIA_HOME.
• Ignore rules hardened for safer open-source sharing.

Highlights

• Add compiled wiki workflows to Memoria with raw source import, generated wiki special pages, query file-back, and durable governance lint findings.
• Extend the CLI and HTTP API so wiki ingest, build, filing, and lint flows are available in both repo mode and installed mode.
• Keep existing memory, bootstrap, no-clone install, adapter runtime, and MCP hybrid flows green under the expanded validation matrix.

Added

• source add and source list for markdown/text raw source import.
• Compiled wiki special pages: knowledge/index.md, knowledge/log.md, and knowledge/overview.md.
• wiki file-query for filing high-value recall results into synthesis or comparison pages.
• wiki lint with durable findings for duplicate, orphan, stale, low-provenance, and source-not-compiled checks.
• CI coverage for wiki ingest, wiki build, query file-back, and wiki lint workflows.

Changed

• Agent guidance, operations docs, release SOP, and skill instructions now treat the compiled wiki as a first-class runtime workflow.
• Compiled wiki flows now refresh alongside session sync and raw source import.

Release Artifact

• Linux x64 tarball: memoria-linux-x64-v1.7.0.tar.gz

Verification

• pnpm run release:docs-check
• pnpm run check
• pnpm run build
• pnpm run release:package
• bash scripts/test-smoke.sh
• bash scripts/test-bootstrap.sh
• bash scripts/test-adapter-runtime.sh
• bash scripts/test-no-clone-install.sh
• bash scripts/test-mcp-e2e.sh
• bash scripts/test-wiki-ingest.sh
• bash scripts/test-wiki-build.sh
• bash scripts/test-wiki-query-fileback.sh
• bash scripts/test-wiki-lint.sh

Highlights

• Add a no-clone Linux x64 installation flow via packaged release artifacts.
• Ship an artifact-based installer and installed-mode bootstrap path.
• Validate repo mode and installed mode in CI, including no-clone end-to-end coverage.