--- name: cognee-memory description: On-demand deep-archive knowledge graph (Cognee). ADDITIVE layer over primary RAG + Zep + memory_search — never replaces them, never in hot path. Failover-safe (no paralysis). Use to recall accumulated arsenal/operational knowledge as an entity/relation graph. ---

Cognee Memory Skill (deep archive graph)

Philosophy (per Dr. Wael 2026-06-29)

• ADDITIVE, never destructive. Primary recall stays: builtin memory_search + the main RAG (ask.drwaelalkishawi.com) + Zep graph. Cognee is a THIRD on-demand layer.
• RED LINE — no paralysis / never stop. Cognee/Gemini/embeddings failure must NEVER block. cognee_query.query() catches all errors and returns degraded=True; the caller keeps using primary RAG + memory_search. Cognee is never in the hot path.
• Privacy split (hard rule):

- NON-sensitive operational/technical archive → cloud cognify (Gemini 3 Flash) + local embeddings. - SENSITIVE (family / real estate / KCPC / personal) → LOCAL ONLY (Ollama cognify + local fastembed), or local vector RAG. Never sent to any cloud.

Stack (verified live 2026-06-29)

• Isolated venv: skills/cognee-memory/venv (uv, Python 3.12). cognee 1.2.2 + transformers + tiktoken + fastembed.
• Embeddings: LOCAL fastembed / intfloat/multilingual-e5-large (1024-dim, Arabic-capable, private, free).
• Cognify LLM (non-sensitive): gemini-3-flash-preview via Google's OpenAI-compatible endpoint (API key, no Vertex). Key = GEMINI_API_KEY.
• Graph + storage: local (kuzu) under skills/cognee-memory/data (+ system).
• Measured POC (8KB): prune 2.7s · embed 27.7s · cognify 62.6s · query 10-26s · quality ~70-80%.

Config

cognee_config.configure(mode="cloud"|"local") — embeddings always local; cloud uses Gemini Flash for cognify (non-sensitive), local uses Ollama qwen (sensitive, slow but private).

Usage

bash
cd skills/cognee-memory
query the non-sensitive arsenal graph (failover-safe JSON out)
./venv/bin/python3 cognee_query.py "what is the golden fallback chain?"
(re)build the non-sensitive archive graph (one-time / when arsenal changes)
. /data/.openclaw/secrets/env.sh && ./venv/bin/python3 build_full.py

When to use / NOT use

• USE: deep relational recall over accumulated arsenal/operational knowledge when memory_search + RAG are not enough.
• NOT: hot path, every turn, or anything time-critical. Always have memory_search + RAG as primary.
• NEVER feed sensitive data through mode="cloud". Sensitive = local tier only.

Notes

• Edges/graph build is one-time per corpus; queries are on-demand.
• If cognee_query returns degraded=True, fall back to memory_search + RAG (no error to user).