v7.2.0 - Evidence-Gated Skill Memory

Mnemosyne v7.2.0 - Evidence-Gated Skill Memory

Mnemosyne v7.2 turns the skill system from a promising prototype into an observable, conservative, evidence-gated loop.

The main idea is simple: an AI agent should not blindly trust a generated skill just because it looks well written. A skill must be tested, receive feedback, survive governance checks, and only then enter the default injection path.

Highlights

• Post-dream skill daemon: full dream cycles can now trigger automatic skill follow-up work.
• Evidence flow: skill usage feedback records success, partial, miss, misleading, and trigger_mismatch outcomes.
• Failure-to-test loop: reproducible failures can become test-prompts.json entries.
• Conservative promotion gate: only low-risk skills with stable success evidence can move toward default injection.
• Dashboard visibility: recent daemon and skill-loop results are visible from the dashboard instead of being hidden in terminal output.
• Open-source packaging: GitHub Pages, CI, issue templates, roadmap, security policy, and visual assets were added for a better public launch.

Why This Matters

Most agent memory systems stop at retrieval. Mnemosyne goes further:

experience -> memory graph -> dream consolidation -> skill embryo -> live evaluation -> feedback -> approved injection

That makes skill growth safer. A generated skill can help, miss, trigger at the wrong time, or even mislead the agent. v7.2 records those outcomes and uses them as governance signals.

What's New

Skill Daemon

The new daemon can run full dream cycles and continue into post-dream skill processing:

skill-daemon.cmd

Default full dream schedule:

03:00, 12:00, 17:00

Skill Feedback Outcomes

memory_skill_feedback now prefers canonical outcomes:

• success
• partial
• miss
• misleading
• trigger_mismatch

These outcomes are more precise than generic helpful/not-helpful ratings and can drive future evaluation sets.

Dashboard Summary

The dashboard now exposes the latest post-dream skill processing summary:

• candidates scanned
• candidates processed
• runner mode
• evolution rounds
• feedback count
• promotion results
• expandable per-candidate details

Public Packaging

This release also improves the public repository:

• GitHub Pages landing page
• CI workflow
• Pages deployment workflow
• issue and PR templates
• roadmap, security policy, and code of conduct
• architecture and dashboard preview SVGs
• social preview asset

Verification

Validated locally with:

python -m py_compile scripts\skill_daemon.py scripts\dashboard\pages\dashboard.py scripts\graph_dream.py scripts\graph_query.py scripts\graph_write.py scripts\graph_audit.py
python scripts\skill_daemon.py --once

The daemon test ran conservatively: candidates that failed live evaluation stayed in needs_revision and were not falsely promoted.

Upgrade Notes

• Run the v7.2 migration if upgrading an existing database.
• Do not commit local runtime databases or llm_config.json.
• If you use GitHub Pages, enable Pages deployment from GitHub Actions.
• Upload assets/social-preview.png as the repository social preview image from GitHub Settings.

Links

• README
• Chinese README
• Open Source Launch Checklist
• v7.2 Development Plan

v5.0.0 — Experience & Memory System

What's New in v5.0

Architecture Shift

From pure memory system → Experience & Memory System. Not just remembering errors, but full lifecycle: successes, failures, decisions, preferences, and patterns.

New Features

• L0/L1/L2 Layered Context — Inspired by OpenViking, 83% token cost reduction
• MCP Server — Zero-dependency stdio JSON-RPC, works with Claude Code / OpenCode / Cursor
• REST API — FastAPI + Swagger UI, 6 endpoints
• Streamlit Dashboard — 4-page visual panel with Kimi-style UI
• Conversation Log Scanner — Auto-extract valuable fragments from opencode conversation logs
• LLM Distillation — Raw fragments → LLM-powered principle extraction
• Dream Log — Full dream history with 13-phase Gantt visualization
• Custom D3.js Graph — Force-directed with zoom/pan, type-colored nodes

Dream Pipeline: 11 → 13 Phases

Added LogScan (Phase 2) and Distill (Phase 11) to the dream pipeline.

Files Changed

32 files changed, 3079 insertions, 148 deletions

v4.1.0 - Plug-in Architecture and Harrier

What's New in v4.1.0

Plug-in Architecture

Every component is now swappable via abstract interfaces:

• AbstractGraphStore → SQLite (default), FAISS, Neo4j, ...
• AbstractEmbedder → Harrier (default), BGE-M3, Qwen, ...
• AbstractTaskRunner → APScheduler (default), Celery, ...

Harrier Embedding Model

Switched from BGE-M3 to Harrier-OSS-v1-0.6b:

• 10x faster model loading (1.2s vs 11s)
• MTEB #1 (2026 benchmark)
• 1024-dim, fully compatible with BGE-M3

11-Phase Dream Pipeline (was 8)

New phases added:

• Phase 1: Snapshot — Pre-dream state capture with safety caps
• Phase 9: LLM Review — Optional REM-style 3-round adaptive review (quick → deep → final)
• Phase 11: Audit — Post-dream health check, bloat detection

Semantic Chain Search

Search results now form coherent semantic chains instead of random hits:

• 55% best-similarity cutoff filters noise
• Graph traversal expansion connects related results

Optional LLM REM Review

• Adaptive 3-round assessment (quick → deep → final)
• Confidence-based actions (high → execute, medium → tentative, low → propose)
• Undo log with before/after snapshots, 7-day auto-purge

Other Improvements

• Principle-based exact classification reinforcement (base_score +0.1)
• Covenant privacy audit — auto-detect and veto sensitive edges
• graph_audit.py — health report + cleanup (template removal, duplicate merge)
• re_embed.py — full re-embedding tool for model swaps
• setup.py — one-command installer
• Windows encoding fix + HF offline helper

Upgrading from v4.0

git pull origin main
python scripts/re_embed.py  # re-embed all nodes with Harrier

Full Changelog: see CHANGELOG.md