- Remove all background REST API transcript dumps (on_pre_compact.py,
capture_compact_summary.py, on_pre_commit.py calls) that were creating
noisy session_state and commit_context memories automatically
- Rewrite pre-compaction prompt to extract individual durable facts
(decisions, learnings, anti-patterns) instead of session summary blobs
- Agent now decides what to store, matching openclaw's triage pattern:
"Most sessions produce zero memory operations. That is correct."
- Fix auto_import.py to search git repo root for CLAUDE.md/AGENTS.md
(was only searching plugin cwd, missing files in parent directory)
- Fix health check session tracker to check stats file directly instead
of depending on CLAUDE_PLUGIN_ROOT env var
- Fix session-end report to print in SessionEnd hook (Stop hook output
doesn't render on /exit)
- Update v2 API endpoint to v3 in git commit capture search
- Keep auto_import.py as only direct REST write path (project profiles,
idempotent, hash-gated)
Root cause: .mcp.json required ${MEM0_API_KEY} shell env var which
OAuth flow never populates — MCP server never started, only auth
stubs were exposed.
Fix: switch .mcp.json to ${user_config.MEM0_API_KEY} so Claude Code
reads the key from plugin userConfig (system keychain). Add
resolve_api_key() fallback chain (MEM0_API_KEY -> CLAUDE_PLUGIN_OPTION_MEM0_API_KEY)
to all hook scripts. Update onboard skill and setup_coding_categories.py
to handle keychain-only users. Add 3 tests for key resolution.
Tour was filtering on metadata.type which missed platform-categorized
memories. Rewritten to use get_memories (list all) + group by platform
categories field. All write hooks now use app_id as top-level entity
param instead of metadata.project_id. Branch stamped on every write.
kartik-mem0