Merge pull request #456 from PR0M3TH3AN/torch-docs-update-13656494316732560712

Update TORCH.md with memory system instructions

TORCH.md

@@ -84,6 +84,31 @@ Use the following folders as lightweight coordination artifacts:
 **Decision note:** Decision, Alternatives considered, Rationale, Consequences/follow-ups.
 **Test log:** Command, Result (pass/fail/warn), Notes (environmental limits, retries, artifacts).
 
+### Long-term Memory
+
+While `src/context` and `src/todo` are ephemeral (per session), TORCH provides a durable memory system for transferring knowledge across runs and agents.
+
+#### Reading Memory
+At the start of a session, the scheduler retrieves relevant memories based on your agent identity and current prompt.
+- **Source:** `.scheduler-memory/latest/<cadence>/memories.md`
+- **Action:** Read this file to learn from past agents (e.g., "The integration test suite is flaky on the login step", "Use `node:test` instead of Jest").
+
+#### Writing Memory
+To save a learning for future agents, create a memory file before finishing your task.
+- **File:** `memory-update.md` (in the repository root).
+- **Content:** Concise, high-value insights.
+- **Mechanism:** The scheduler will automatically ingest this file after your run completes.
+
+**Good Memories:**
+- "Fix for `EADDRINUSE` in tests: ensure server.close() is called in `after()`."
+- "Project preference: Always use named exports for components."
+- "Constraint: Do not upgrade `dependency-x` past v2.0 due to breaking change Y."
+
+**Bad Memories:**
+- "I updated file X." (Use git history for this).
+- "Running tests..." (Use task logs).
+- Large code blocks or entire file dumps.
+
 ### Knowledge-sharing protocol
 
 #### What belongs where