v5.0.0 — Talewell, plugin-first long-term memory

Talewell — plugin-first long-term memory for every agent platform

This release marks an architecture change, not a regression. The previous line (4.1.0 and earlier) was OpenClaw Auto-Dream, a prompt-based memory skill. Talewell is that same idea rebuilt as a portable, plugin-first framework — because every major agent platform is plugin-shaped now.

Added

L1 kernel — plain-Markdown memory repo, zero npm dependencies:

• records with a portable frontmatter contract (facts / entities / episodes / procedures)
• git-backed: every write is a commit, every fact is diff-able and rollback-able
• deterministic recall — relevance × recency decay × importance, computed locally, no API key, no network
• provenance gate — untrusted and system origins are structurally refused
• corrections that stick: superseding retires the old value from recall and injection, while git keeps the history
• [[id]] link graph with stable IDs and aliases

Adapters, by host capability tier:

• adapters/mcp/ — zero-dependency MCP stdio server. Six tools: memory_recall, memory_index, memory_get, memory_remember, memory_forget, memory_graph.
• adapters/openclaw/ — companion plugin with before_prompt_build injection, so the agent starts every turn already knowing what it knows.
• adapters/file/ — universal fallback: an idempotent managed block in AGENTS.md / CLAUDE.md.
• adapters/acp/ — context injection and conservative candidate harvest for host-driven harnesses.

CLI — init, remember, recall, correct, forget, graph, index, health, log, diff, rollback, bundle, consolidate, serve-mcp, hosts, install, uninstall.

Docs — README.md and README.zh-CN.md rewritten; docs/ARCHITECTURE.md added; docs/HOSTS.md added with per-platform setup.

Tests

28 tests, all passing: round-trip serialisation, commit-per-write, the taint gate, deterministic CJK recall, supersession, git-backed forget, graph traversal, rollback, gated consolidation, and the host-file adapter.

Preserved

The original skill and its references live under legacy-skill/ and still install and work on OpenClaw. ClawHub's openclaw-auto-dream name is unchanged.

Non-negotiables

• No vector search in the kernel. The baseline runs offline, with no API key.
• No silent writes. Untrusted content cannot reach durable memory.
• No chat-time curation. Writes go through consolidation, not the reply path.

Full diff: v4.1.0...v5.0.0

v4.1.0 — Cross-Platform Tooling

What's Changed

Cross-platform tooling — dream prompts no longer depend on Unix shell commands. Massive thank-you to @labsclaw for the contribution (PR #4).

Added / Changed

• Dream prompts (dream-prompt-lite.md, first-dream-prompt.md) now use OpenClaw native tools — read to inspect files, edit/write to mutate — instead of ls, wc -l, grep -c, cat, python -c
• File discovery and metric counting happen in the agent's reasoning rather than via shelling out
• Fixes silent failures on Windows, where Unix shell commands are unavailable

Files Changed (bced566)

• references/dream-prompt-lite.md (+29 / -24)
• references/first-dream-prompt.md (+34 / -30)

Docs

• Version badges bumped to 4.1 across all 9 READMEs (f6e0be2)
• Release notes added

Full diff: v4.0.0...v4.1.0

v4.0.0 — Intelligent Dream UX

🌟 Intelligent Dream UX (2026-03-30)

Dream cycles that feel alive.

• Skip-with-Recall: When no new content, surfaces an old memory reminder + streak count instead of empty skip
• Cumulative Growth Metrics: Notifications show "142 → 145 entries (+2.1%)" + dream streak
• Stale Thread Detection: Scans Open Threads for items >14 days untouched, surfaces top 3 in notification
• Milestone Celebrations: Triggers at 1st/7th/30th dream, 100/200/500 entries
• Weekly Summary: On Sundays, generates week-over-week growth and biggest memories
• Auto-Refresh Dashboard: Regenerates dashboard.html each dream cycle
• Emotional Anchor (First Dream): Personalized reflection paragraph for deeper connection

v3.6.0 — First Dream AHA Moment

✨ First Dream AHA Moment (2026-03-30)

• Instant First Dream: Post-install scan runs immediately — no 24h wait
• Before/After Comparison: Shows memory state change in real-time
• Empty instance handling: Works even with zero daily logs
• New reference: first-dream-prompt.md for post-install experience

v3.5.0 — English Rewrite

🌐 English Rewrite (2026-03-30)

• Full SKILL.md rewritten to English for international audience

v3.4.0 — Dashboard Redesign

🎨 Dashboard Redesign (2026-03-29)

• Chinese-native dashboard: Complete UI redesign with mobile responsive layout
• Placeholder system: Template uses __DREAM_DATA_PLACEHOLDER__ for dynamic data injection

v3.3.0 — Framework Fix

🔧 Framework Fix (2026-03-29)

First successful dream cycle!

• Initialized memory infrastructure: Created all missing files (index.json, dream-log.md, procedures.md, archive.md)
• Simplified dream prompt: dream-prompt-lite.md rewritten to 1188 bytes for reliable cold-start
• Slimmed SKILL.md: Removed overengineering, focused on practical execution
• Increased timeout: 300s → 600s for complex first runs
• 3-day scan range: Daily incremental runs process only recent unconsolidated logs

v3.2.1 — Security & LCM Fix

🔒 Security & LCM Fix (2026-03-29)

• LCM detect-only mode: No longer auto-installs plugins; prompts user instead
• ClawHub security flag resolved: Removed auto-install commands that triggered scanner
• Sensitive data policy: Renamed from "Secrets policy" to avoid false positives

v3.2.0 — Performance Optimization

⚡ Performance Optimization (2026-03-29)

~90% token savings on idle days

• Smart Skip: Dream cycle checks for unconsolidated daily logs first. If nothing new → instant exit (~2K tokens vs ~150K)
• Lite Prompt: Self-contained 97-line prompt replaces 928 lines across 3 files
• Faster Execution: Cron reads 1 file instead of 3. Timeout reduced 600s → 300s
• Full prompt preserved: dream-prompt.md (311 lines) kept for manual deep runs

v3.1.1 — Language & Stability

🌍 Language & Stability (2026-03-28)

• User language support: Dream notifications/reports auto-detect language from USER.md
• Notification redesign: Uses cron delivery:announce instead of direct message tool calls
• 7 bug fixes: Scoring normalization, health formula, dashboard compatibility, reference integrity