TT-Wang/cortex-plugin

Hotfix for v0.10.0. Post-ship forge code review found six bugs that the test suite missed (tests were mocking around the real data paths).

Correctness

• memory_timeline read mem.get("created") but the parser stores the field as created_at. Every timestamp was 0.0 so the "Before" section was always empty. Fixed all six call sites.
• pre-tool-use.sh drained stdin into $INPUT then tried to json.load(sys.stdin) in a heredoc — stdin was already empty, hook silently fell through to empty context. MEMEM_PRETOOL_GATING=1 was a no-op for every user. Fixed by writing stdin to a tempfile and passing the path via argv.
• classify_layer read mem.get("tags") but memories store their tag list under domain_tags. Tag-based L0/L1 classification never fired. Same bug in _format_full_memory (memory_get display). Both fixed to domain_tags or tags.

Performance

• O(N²) vault rescans — classify_layer(mem, _obsidian_memories()) was inside the mining loop. Hoisted the snapshot once per session.
• --compact-index was unbounded — 290 KB / ~72K tokens injected at every session start for a 1895-memory vault. Added --limit N flag (default 500, override via MEMEM_COMPACT_INDEX_LIMIT), ranked by importance + recency. Drops to ~20K tokens.

Graph traversal

• FTS path silently did 2-hop (_search_memories expanded, then memory_search expanded again). Added expand_links=False parameter; memory_search now does a single-hop expansion. memory_recall keeps the old behavior for backward compat.

MCP schema

• memory_timeline min_length=4 but _find_memory requires 8-char prefixes. Fixed to 8.
• memory_timeline scope_id was accepted but never used. Fixed to pass through to _obsidian_memories.

Known trade-off

PreToolUse hook has ~5s cold-start latency because it spawns a fresh python -m memem.server --recall subprocess that imports the full index. That's why it's opt-in via MEMEM_PRETOOL_GATING=1. A future release will query search.db directly via a lightweight helper.

73/73 tests still green, ruff clean. Upgrade: /plugin update memem.