Memory Skill

Operates on layers 2 and 3 of the AIGORA memory stack. Layer 1 (native auto-memory) is managed by Claude Code itself; see .claude/rules/memory-protocol.md for the full architecture.

When to use this skill

• The user references something from a past session that is not in the native auto-memory topic files → search layer 3.
• The user asks "what have you got on me?" or wants an export → list layer 3.
• You want to manually add a fact you judge important enough → add to layer 3.
• You want to append a chronological event to today's log → write to layer 2.
• A skill (weekly-review, meeting-prep, business-setup) needs to persist project facts → use the Write tool against layer 1 topic files directly. Do NOT use this skill for layer 1 writes.

Layer 3 operations (mem0 + Qdrant, optional)

Search (recommended: smart_search)

python3 .claude/skills/memory/scripts/smart_search.py --query "topic" --limit 5

Hybrid retrieval: BM25 keyword (30%) + vector similarity (70%) + temporal decay + MMR diversity. First run on a fresh install requires --rebuild-index to populate the FTS5 keyword index.

Search (basic vector only)

python3 .claude/skills/memory/scripts/mem0_search.py --query "topic" --limit 10

Rebuild keyword index

python3 .claude/skills/memory/scripts/smart_search.py --rebuild-index

Repopulates FTS5 from the history DB. Run after installation or to fix drift.

Add a fact manually

python3 .claude/skills/memory/scripts/mem0_add.py --content "User prefers kSuite over Google Workspace"

Add from conversation messages

python3 .claude/skills/memory/scripts/mem0_add.py --messages '[{"role":"user","content":"I switched to ClickUp"}]'

List all memories

python3 .claude/skills/memory/scripts/mem0_list.py --limit 50

Delete a memory

python3 .claude/skills/memory/scripts/mem0_delete.py --memory-id "abc123"
python3 .claude/skills/memory/scripts/mem0_delete.py --all --confirm

Layer 2 operation (daily logs)

Append to today's session log

python3 .claude/skills/memory/scripts/daily_log.py --content "Completed memory system overhaul" --type event

Auto-capture (background, no action needed)

Stop hook auto_capture.py runs after every response cycle. It reads new transcript messages, feeds them to mem0 for fact extraction and dedup against the vector store, and appends a session summary to today's daily log. Logs at data/auto_capture.log.

Config

• mem0 config: .claude/skills/memory/references/mem0_config.yaml
• Default (100% local, open source): LLM qwen3:4b and embeddings nomic-embed-text served by Ollama on this computer; vectors in Qdrant embedded mode (data/qdrant/, no server).
• Alternative LLM: Infomaniak AI Tools (OpenAI-compatible, open-source models hosted in Switzerland), see the commented block at the bottom of the config.
• History DB: data/mem0_history.db (SQLite, auto-managed)
• Capture markers: data/capture_markers/ (tracks transcript position per session)

Security

• sanitize_text() in mem0_client.py strips secrets (API keys, tokens, JWTs, connection strings) before any processing.
• With the default local stack, nothing leaves the computer: extraction, embeddings and storage are local.
• If you switch the LLM to Infomaniak AI Tools, conversation snippets are sent to Infomaniak (Switzerland).
• Local files (SQLite history DB, Qdrant folder, daily logs) are plaintext on disk; .gitignore excludes them.

Known issues

1. Small local models can occasionally return invalid JSON on heavy markdown. prepare_messages() strips most of it; failed batches are skipped, the transcript is preserved. A larger model (qwen3:8b) is more reliable if the machine allows.
2. First run on a long conversation is slow on CPU. Incremental runs (2-4 messages) are fast.
3. Layer 3 is off by default. Layers 1 and 2 (native memory + daily logs) always work with no install.