59 lines
2.9 KiB
Markdown
59 lines
2.9 KiB
Markdown
# Agentic-OS memory layer + Graphify — honest status
|
|
|
|
## Can Graphify BE the memory layer?
|
|
**Partially, and not as a replacement for the write path.**
|
|
|
|
What Graphify is: a tool that turns a *codebase* into a knowledge graph
|
|
(nodes = code symbols/modules, edges = relationships) and lets you
|
|
visualize / query it. It is a **read/visualize/query layer over your
|
|
files** — it does NOT store or write memory. The OS's actual
|
|
memory write path stays `brain/memory.md`, `append_audit()`,
|
|
`data/*.json`, etc.
|
|
|
|
## What we verified (2026-07-25)
|
|
- `graphify . --code-only` over the whole project → **321 nodes, 517
|
|
edges, 41 communities** (graphify-out/graph.json + graph.html).
|
|
This graphs the *code* (server.py, skills' .py, etc.).
|
|
- `graphify brain skills --code-only` → **0 nodes**. The memory
|
|
folders are almost entirely **Markdown**, which `--code-only`
|
|
skips. So the memory *content* is invisible to a code-only build.
|
|
- To graph the *memory docs* (markdown notes), Graphify needs an
|
|
LLM key for semantic extraction (GEMINI_API_KEY / GOOGLE_API_KEY
|
|
/ ANTHROPIC_API_KEY / OPENAI_API_KEY / DEEPSEEK_API_KEY /
|
|
MOONSHOT_API_KEY). With a key:
|
|
`graphify brain skills` (no --code-only) will build a graph that
|
|
includes your notes as nodes. Without a key it errors out.
|
|
|
|
## Recommended setup (works today, upgrades with a key)
|
|
1. Build the code graph (no key needed):
|
|
cd ~/agentic-os && graphify . --code-only
|
|
→ live in the dashboard at http://localhost:8080/graphify-out/graph.html
|
|
and the "Agent Time" page's "Open interactive graph" link.
|
|
2. (Optional, with a key) Build the memory-inclusive graph:
|
|
export GEMINI_API_KEY=...
|
|
graphify brain skills data # docs + code, semantic
|
|
This adds your memory notes as queryable nodes.
|
|
3. Expose it to the OS agents as a query tool (MCP or HTTP):
|
|
# HTTP (team/shared):
|
|
python -m graphify.serve graphify-out/graph.json --transport http --port 8081
|
|
# or stdio MCP for a single agent:
|
|
python -m graphify.serve graphify-out/graph.json
|
|
Then point the agent/MCP client at it. Agents can ASK the graph
|
|
("what connects memory-consolidation to audit?", "show the
|
|
agentic-os data flow") — a retrieval aid, not a memory writer.
|
|
|
|
## Bottom line
|
|
- ✅ Use Graphify as a **visualization + semantic-query layer** for the
|
|
OS's memory/code. Great for "what's in my brain and how does it
|
|
connect."
|
|
- ❌ Do NOT treat it as the memory store. Writing memory still goes
|
|
through the existing `brain/` + `audit` + `data/` files. Graphify
|
|
reads those; it never writes them.
|
|
- ⚠️ Memory-note graphing requires an LLM key; code graphing does not.
|
|
|
|
## Files
|
|
- graphify-out/ — built code graph (graph.json, graph.html, GRAPH_REPORT.md)
|
|
- scripts/analyze_agent_time.py — agent-time estimator (the dashboard's data source)
|
|
- dashboard/.../agent-time.js — integrated "Agent Time" page (port 8080)
|
|
- dashboard/serve_monitor.py — standalone monitor (kept, unused by integration)
|