Improve agent protocol: peek-then-drill, concrete write triggers, multi-point recall

- Rewrite README workflow: two-phase recall (peek brief → drill get), recall
  trigger table (6 moments), write trigger table (6 types), end-of-task writes
- Update MCP tool descriptions to reinforce peek-then-drill pattern:
  memory_recall emphasizes format=brief first, memory_get is the drill step,
  memory_add lists concrete trigger points
- Address real agent behavior problems: agents reading body on broad queries,
  only recalling at session start, not knowing when to write memories

Author: bburda

README.md

@@ -209,17 +209,66 @@ Tags use `prefix:value` format for consistent discovery:
 
 ### Recommended Workflow
 
-1. **RECALL before working** — always check existing knowledge before starting a task
-2. **ADD after learning** — record discoveries, decisions, and observations immediately
-3. **SUPERSEDE, don't edit** — when knowledge changes, add a new memory and deprecate the old one
-4. **CHECK STALENESS periodically** — use `memory stale` to maintain knowledge quality
+#### 1. READ — Peek then Drill (two-phase recall)
+
+Always use a **two-phase** approach. Never go straight to body text on broad queries.
+
+**Phase A — Peek** (scan titles, zero body text):
+```bash
+memory recall --tag topic:gateway --format brief --expand 0
+```
+Returns `[ID] Title (confidence)` one-liners. Minimal tokens. Do this FIRST.
+
+**Phase B — Drill** (read full body of specific memories):
+```bash
+memory get DEC_handler_context_pattern
+```
+Only after peeking — pick the 2-3 most relevant IDs and `get` them individually.
+
+**When to recall** — recall is NOT just a session-start ritual. Recall at each of these moments:
+
+| Trigger | What to recall |
+|---------|---------------|
+| Session start | `recall --format brief --limit 20 --sort newest` |
+| New task or topic | `recall --tag topic:<X> --format brief` |
+| Entering unfamiliar code | `recall --tag repo:<X> --type fact --format brief` |
+| Before a design decision | `recall --tag topic:<X> --type dec` |
+| When stuck or debugging | `recall <error keywords>` |
+| Before implementing a pattern | `recall --tag intent:coding-style --type pref` |
+
+#### 2. WRITE — Record at specific trigger points
+
+Recording memories is NOT optional. Write at these concrete moments:
+
+| Trigger | Type | Example |
+|---------|------|---------|
+| Chose approach A over B | `dec` | "Use tl::expected over exceptions" |
+| Fixed a non-obvious bug | `mem` | "EntityCache race condition fix" |
+| Discovered undocumented API | `fact` | "Routes match in registration order" |
+| User stated a preference | `pref` | "Prefer Zustand over Redux" |
+| Identified a risk | `risk` | "JWT secret hardcoded in tests" |
+| Question remains unanswered | `q` | "Should synthetic components expose operations?" |
+
+**End-of-task writes**: summarize architecture learned (`fact`), record conventions (`pref`), note anything a future agent needs (`mem`), capture unfinished goals (`goal`).
+
+**Write quality rules**:
+- `--tags` is mandatory — without tags, the memory is unfindable
+- `--body` must be self-contained with file paths and concrete details
+- Use `--rebuild` flag to make new memories immediately searchable
+
+#### 3. SUPERSEDE, don't edit
+
+When knowledge changes, add a new entry with `--supersedes OLD_ID` and deprecate the old one.
+
+#### 4. CHECK STALENESS periodically
+
+Run `memory stale` at the start of long sessions to keep the graph accurate.
 
 ### Context Window Optimization
 
-- Default `recall` omits body text — use `memory get <ID>` only when you need details
-- Use `--format brief` for initial scanning, then drill into specific IDs
-- Use `--limit 10` when exploring broad topics
-- Use `--expand 0` to skip graph expansion for exact matches only
+- `recall` omits body by default — this is intentional, not a limitation
+- **Peek** with `--format brief` → **drill** with `get <ID>` — this is the core pattern
+- Use `--limit 10` and `--expand 0` when exploring broad topics
 - Use `--tag` filters to narrow results instead of free-text
 - Use `memory tags` to discover available tag prefixes before filtering
 

src/ai_memory_protocol/mcp_server.py

@@ -67,7 +67,8 @@ def create_mcp_server(name: str = "ai-memory-protocol") -> Server:
         description=(
             "Search memories by free text query and/or tags. "
             "Returns matching memories formatted for context windows. "
-            "Use this FIRST to check existing knowledge before starting work."
+            "Use format='brief' FIRST to peek at titles, then memory_get to drill into specific IDs. "
+            "Recall at EVERY topic transition: new task, unfamiliar code, before decisions, when stuck — not just session start."
         ),
         inputSchema={
             "type": "object",
@@ -123,8 +124,9 @@ def create_mcp_server(name: str = "ai-memory-protocol") -> Server:
     Tool(
         name="memory_get",
         description=(
-            "Get full details of a specific memory by ID. "
-            "Always shows body text. Use after recall to drill into a specific memory."
+            "Get full details of a specific memory by ID — the DRILL step. "
+            "Always shows body text. Use AFTER memory_recall(format='brief') to read "
+            "the 2-3 most relevant memories. Never skip the peek step."
         ),
         inputSchema={
             "type": "object",
@@ -140,9 +142,12 @@ def create_mcp_server(name: str = "ai-memory-protocol") -> Server:
     Tool(
         name="memory_add",
         description=(
-            "Record a new memory. Use when you discover, decide, or observe something important. "
-            "Always include tags for discoverability. Always include body with enough context "
-            "that a future agent can act on it without re-reading source files."
+            "Record a new memory. You MUST use this at specific trigger points: "
+            "chose approach A over B → dec; fixed non-obvious bug → mem; "
+            "discovered undocumented pattern → fact; user stated preference → pref; "
+            "identified risk → risk; question unanswered → q. "
+            "Also batch-write at end of task: architecture learned → fact, conventions → pref. "
+            "Tags are mandatory. Body must be self-contained with file paths and concrete details."
         ),
         inputSchema={
             "type": "object",