v3.5.0 — Studio: Connect Agents + systematic bug fixes

Highlights

🔌 Connect Agents — one-click MCP setup

Studio now ships a Connect Agents view (/connect, keyboard G C) that detects installed MCP-capable AI clients and wires Mnemion into each one with a single click.

Supported: Claude Code (user + project), Claude Desktop, OpenAI Codex CLI (TOML), Google Gemini CLI, Cursor, Codeium Windsurf, Zed. Legacy mempalace entries are detected and auto-replaced. Every install writes a timestamped backup to .mnemion_backups/. The installed command uses the absolute path of Studio's own Python interpreter — no PATH surprises.

Any unlisted client (OpenClaw, Nemoclaw, Hermes, Cline, custom agents…) can connect via the universal JSON snippet in the view.

🔴 Critical Studio bug fixes

• Timestamp field mismatch — readers queried timestamp; writers save filed_at. Every "Created" row, Recently Added sort, and agent last-seen was silently broken.
• DrawerCreateModal didn't navigate — read data.id when backend returns drawer_id.
• Empty search previews — hybrid_searcher returns text, frontend rendered content; now mapped at the API boundary.
• Vault export crashed on click — referenced undefined _col; now uses _get_collection() and streams via FileResponse with a temp file.
• CORS blocked non-5173 ports — widened to any localhost port plus file://.
• CommandPalette showed literal "undefined" — operator-precedence bug in fallback.
• Hardcoded port 5173 in Settings, Electron dev mode, and start.bat — all now dynamic.

🟢 Studio features

• Graph hover highlight — Obsidian-style: hover a node, non-neighbours dim, adjacent edges brighten.
• Dashboard recent drawers + quick-capture button.
• Search wing: / room: inline operators; contested trust badges on results.
• <ErrorBoundary> wraps the router <Outlet />.
• New endpoint: GET /api/drawers/recent.

🟡 Cleanup

• Dead RightSidebar.tsx (162 lines, never imported) removed.
• Dead /ws WebSocket stub + broadcast() removed.
• hooks/mnemion_save_hook.py no longer hardcodes ~/projects/mnemion.
• Shortcut modal reflects only wired shortcuts; added G C → Connect Agents.

Upgrade

pip install --upgrade mnemion
uvicorn studio.backend.main:app --port 7891
cd studio/frontend && npm install && npm run dev

Open http://localhost:5173 → Connect Agents → click Install on each AI client.

Compatibility

Drop-in upgrade from v3.4.x. Legacy mempalace MCP entries are auto-replaced on install. All 120 Python tests pass.

---

🤖 Co-authored with Claude Code

Mnemion v3.4.7 — Surgical Hardening & Ghost Purge

Mnemion Surgical Hardening & Ghost Purge (v3.4.7)

This release marks a final transition to a pure, production-ready Mnemion core by purging technical debt and legacy ghosts.

🛠️ Key Improvements

• Legacy Ghost Purge: Excised all remaining hardcoded references to mempalace, palace, and MemPal from config systems, miners, and benchmarks.
• Dependency Hardening: Added explicit logger.warning notifications when orch is missing but latent space grooming (LeWM) or context prediction (JEPA) is configured. No more silent feature degradation.
• Onboarding Cleanup: Removed nonexistent palace_facts.py references from generated metadata.
• Benchmark Synchronization: Realigned all internal benchmarks (ConvoMem, LongMemEval, LoCoMo) to the native repository and Mnemion namespace.
• Version Manifest Parity: Fully synchronized versions across .molthub/project.md, README.md, .claude-plugin, and .codex-plugin.

🧹 Maintenance

• Removed MempalaceConfig compatibility alias from config.py.
• Stripped legacy ~/.mempalace migration fallbacks for architectural purity.
• Updated project version to v3.4.7.

v3.4.1 — Mnemion Anaktoron Hardening & Truth Audit

Finalized Anaktoron-native architecture, synchronized FTS5/ChromaDB/Trust pipelines, and integrated LEWM predictive latent trajectory.

v3.3.5 — Streaming restore, O(batch) peak memory

Root cause of SIGKILL

restore called json.load() on the full export before writing anything. A 58 MB file materialises as ~500 MB–1 GB of Python objects (dict/string overhead). Combined with ChromaDB loading its sentence-transformer model (~90 MB) and running embedding inference, the OOM killer was triggered consistently at ~950/33 433 drawers — before even 3% of the archive was written.

Reducing batch size didn't help because the full list was already in RAM.

What changed

_stream_json_array(filepath)

Yields one drawer object at a time using json.JSONDecoder.raw_decode() with a 512 KB rolling file buffer. At any point during restore, only the current batch is in memory — regardless of archive size.

_count_json_objects(filepath)

Fast byte scan (b'"id":') that counts drawers in ~20 ms without JSON parsing. Used at startup so % progress still works with the streaming path.

cmd_restore

Now feeds _stream_json_array() into the batch loop. The full export never exists as a Python list.

Memory profile (33 k drawer archive)

Version	Peak RAM during restore
v3.3.2 (batch GC)	~500–1 000 MB (full list in RAM)
v3.3.5 (streaming)	~90–150 MB (transformer + one batch)

v3.3.2 — Restore OOM fix

What's fixed

SIGKILL during large archive restore

Restoring a 33k-drawer, 58MB export was dying with SIGKILL (OOM killer) because:

• ChromaDB embeds every document on write using a sentence transformer model
• Batch size of 500 × ~22k char average = ~11MB of text per embedding round
• The full JSON array stayed in memory throughout

Changes

• Default batch size 500 → 50 — bounds the embedding memory per round
• --batch-size flag — mnemion restore archive/drawers_export.json --batch-size 20 for very tight hosts
• Per-batch GC — each batch is cleared from the in-memory list and gc.collect() is called after every ChromaDB write
• flush=True on all print() — progress is now visible before any OOM event
• Progress shows % + drawer count — [35%] 11700/33433 ...
• README updated — all mnemion mine archive/... references replaced with mnemion restore, OOM note added with --batch-size workaround

v3.3.0 — restore command + collection name fixes

What's new

mnemion restore — new command for importing a JSON export

The mnemion mine archive/drawers_export.json path documented in the README was broken — mine expects a directory, not a JSON file, and would crash with No mnemion.yaml found. This release adds the actual command:

```bash

New machine setup from shared export

mnemion restore archive/drawers_export.json

Add to an existing palace without wiping it

mnemion restore archive/drawers_export.json --merge

Replace an existing palace entirely

mnemion restore archive/drawers_export.json --replace
```

Collection name resolution fixed across all commands

Several commands (search, wake-up, status, repair, compress, and the miner) hardcoded "mnemion_drawers" instead of reading collection_name from ~/.mnemion/config.json. This caused silent failures for palaces created with the legacy "mempalace_drawers" collection name.

All read/write paths now resolve the collection name from MempalaceConfig().collection_name.

Files changed

• cli.py — restore command, repair and compress collection name fix
• searcher.py — search() and search_memories() collection name fix
• layers.py — Layer1, Layer2, Layer3, MemoryStack.status() collection name fix
• miner.py — get_collection() and status() collection name fix
• convo_miner.py — get_collection() collection name fix

v3.2.23 — Multi-agent palace sync

Multi-Agent Palace Sync

Multiple agents on different machines can now share the same memory palace automatically — no human coordination, no corrupted JSON, no silent failures.

New files

sync/merge_exports.py
Pure-Python merge utility. Produces a clean union of two drawers_export.json files without git merge markers. When the same drawer ID exists in both, the newer filed_at timestamp wins (remote wins on tie). Handles first-sync (missing file) gracefully.

sync/SyncMemories.sh
Bash implementation of the sync algorithm for Linux/macOS agents (openclaw, headless servers).

Rewritten: sync/SyncMemories.ps1

Before	After
Blind push — fails silently if remote is ahead	git fetch before every push
No merge — git conflict markers corrupt the JSON	merge_exports.py produces a valid merged JSON
No retry — one rejection and it gave up	Up to 5 retries with 2–9 s random jitter
git push origin master hardcoded	Branch auto-detected; overridable via MNEMION_BRANCH
No agent identity in commits	MNEMION_AGENT_ID stamped in every commit
No concurrency guard	Lock file; stale locks > 10 min auto-cleared

Environment variables

Variable	Default	Description
MNEMION_AGENT_ID	hostname	Name in commit messages
MNEMION_BRANCH	auto-detected	Branch to sync
MNEMION_REPO_DIR	~/.mnemion	Memory git repo path
MNEMION_SOURCE_DIR	auto-detected	Mnemion package path
MNEMION_PYTHON	python3	Python binary (Linux/Mac)

Known v1 limitation

Drawer deletions do not propagate. A drawer deleted on one agent will be re-added after the next merge from another agent that still has it. Tombstone propagation is planned for a future version.

v3.2.22 — Entity detection quality, search ranking, Makefile

What's fixed

Entity detection false positives

• ~120 new stopwords added: status adjectives (current, verified, pending, active…), common tech/business nouns (stage, trust, hybrid, call, notes, auto…), adjective-nouns that appear capitalised in project docs (lexical, semantic, abstract…). Directly addresses the reported noisy candidates.
• Frequency threshold raised 3 → 5: words that appear fewer than 5 times no longer become candidates, reducing sentence-start capitalisation noise.
• Uncertain list filtered: zero-signal entries with confidence < 0.3 are now dropped before presentation; uncertain cap tightened 8 → 6.

Search ranking (conversational queries)

_fts_search previously ran only a strict phrase match (entire query in double-quotes). For conversational queries like "Which APIs are next after Shopify and CJ?" the phrase never matched anything, leaving ranking entirely to vector search — which pulled broad overview docs (USER.md, TOOLS.md) ahead of the directly relevant ads.md.

Now runs two passes: phrase search (preserved for exact-phrase priority) + keyword search (stop-words stripped, AND-of-terms). Both candidate sets are merged before RRF fusion.

Pytest environment binding

New top-level Makefile with install, test, test-fast, lint, format, clean targets. All test targets invoke $(VENV_PY) -m pytest, so pytest always runs in the project venv.

Fix for: ConftestImportFailure: No module named 'chromadb' — caused by running the system-level pytest binary instead of the venv one.

make test        # full suite
make test-fast   # skip benchmark/slow/stress marks

Tag alignment

Tags v3.2.19, v3.2.20, v3.2.21 now point to their correct version-bump commits. Previously v3.2.19 was misplaced at HEAD (the 3.2.21 commit), causing the reported version inconsistency.

v3.2.21 — version bump

Version bump only. No code changes.

v3.2.20 — version bump

Version bump only. No code changes.