Added tl draw mcp server and skill. Added the fifth visual memory layer as well, to be worked with tl draw only at localhost 3000

Author: Siddharth11Roy

.agent/AGENTS.md

@@ -11,6 +11,8 @@ same memory, skills, and protocols.
 - `memory/semantic/DECISIONS.md` — past architectural choices
 - `memory/semantic/LESSONS.md` — distilled patterns (rendered from `lessons.jsonl`)
 - `memory/episodic/AGENT_LEARNINGS.jsonl` — raw experience log (top-k by salience)
+- `memory/visual/INDEX.md` — durable canvas snapshots from the tldraw skill
+  (fifth layer, opt-in via the `tldraw` feature flag)
 
 ## Review Queue (host-agent responsibility)
 

.agent/memory/visual/INDEX.md

@@ -0,0 +1,8 @@
+# Visual memory index
+
+Rendered from `snapshots.jsonl`. Do not hand-edit entries — re-render by
+calling `visual_memory.py snapshot|archive` or the module's `_render_index`.
+
+## Active
+
+_(no snapshots yet)_

.agent/memory/visual/README.md

@@ -0,0 +1,68 @@
+# Visual memory
+
+The fifth memory layer. Every other layer is text; this one is pictures.
+
+```
+working/   scratchpad for the current task
+episodic/  raw experience log
+semantic/  distilled lessons
+personal/  user preferences
+visual/    drawings the agent and user share   <-- you are here
+```
+
+Visual memory lives as snapshots of a live tldraw canvas (see the
+`tldraw` skill). The canvas itself is ephemeral — the browser tab closes,
+the drawing is gone. Snapshots make a specific canvas state durable and
+recallable.
+
+## Shape
+
+```
+visual/
+  README.md             this file
+  snapshots.jsonl       source of truth: one metadata record per snapshot
+  INDEX.md              rendered view of snapshots.jsonl (human-readable)
+  snapshots/
+    <id>.json           full canvas state for snapshot <id>
+    archive/            archived (never deleted) snapshots
+  visual_memory.py      CRUD module + CLI
+```
+
+`snapshots.jsonl` is the source of truth. `INDEX.md` is re-rendered from
+it; do not hand-edit. This mirrors the `lessons.jsonl` / `LESSONS.md`
+pattern in `semantic/`.
+
+## CLI
+
+```bash
+# capture the current canvas (shapes JSON on stdin)
+<get_canvas output> | python3 .agent/memory/visual/visual_memory.py \
+    snapshot --label "auth-flow-v1" --tags architecture,auth \
+    --note "agreed login + refresh flow"
+
+# list stored snapshots
+python3 .agent/memory/visual/visual_memory.py list [--tag architecture]
+
+# load a snapshot's full shape data
+python3 .agent/memory/visual/visual_memory.py load <id>
+
+# archive a stale snapshot (moves to snapshots/archive/, never deletes)
+python3 .agent/memory/visual/visual_memory.py archive <id>
+
+# show layer status
+python3 .agent/memory/visual/visual_memory.py status
+```
+
+The module is also importable:
+
+```python
+from visual_memory import snapshot, list_snapshots, load_snapshot, archive_snapshot
+```
+
+## Rules
+
+- Append-only. `archive` moves a snapshot into `snapshots/archive/`. Nothing
+  is ever deleted from disk.
+- Ids are time-sortable so `ls snapshots/` reads chronologically.
+- Snapshots are self-contained: a single `<id>.json` carries everything
+  needed to restore the canvas. `snapshots.jsonl` is an index over them.