Refactor memory action management and enhance document event handling

- Introduced `apply_actions_to_text` function to streamline action application on memory text.
- Updated `MemoryActionManager` to utilize the new function for applying actions.
- Enhanced `CollectionMemoryManager` to read memory text safely with `read_memory_text_or_empty`.
- Added `apply_update_to_text` function for handling memory updates based on proposals.
- Expanded `WebhookEventModel` to include `DocumentEventModel` for better event handling.
- Implemented document event handling in `processor_request.py` to refresh memory on document updates and deletes.
- Updated tests to verify memory handling on document events and ensure proper memory state management.
- Modified `CollectionWorkspaceManager` to bootstrap only system prompts until memory is needed, improving initialization efficiency.

Author: visualDust

.github/workflows/build-and-publish-container-image.yml

@@ -4,6 +4,9 @@ on:
   push:
     tags:
       - 'v*'
+  release:
+    types:
+      - published
   workflow_dispatch:
 
 env:
@@ -54,7 +57,7 @@ jobs:
           cache-to: type=gha,mode=max
 
       - name: Create GitHub Release
-        if: startsWith(github.ref, 'refs/tags/v')
+        if: github.event_name == 'push' && startsWith(github.ref, 'refs/tags/v')
         uses: softprops/action-gh-release@v2
         with:
           generate_release_notes: true

config.example.yaml

@@ -119,6 +119,8 @@ features:
 runtime:
   # If true, plan and simulate side effects without actually mutating Outline state.
   dry_run: false
+  # Title of the remote Outline document used for collection memory.
+  collection_memory_document_title: MEMORY
   # Maximum number of plan/replan rounds for one request.
   tool_execution_max_rounds: 10
   # Maximum number of steps the planner may propose in one round.

docker/config.yaml.example

@@ -59,6 +59,7 @@ features:
 
 runtime:
   dry_run: false
+  collection_memory_document_title: MEMORY
   tool_execution_max_rounds: 10
   tool_execution_max_steps: 6
   tool_execution_chunk_size: 2

docs/config.md

@@ -93,6 +93,7 @@ features:
 
 runtime:
   dry_run: false
+  collection_memory_document_title: MEMORY
   tool_execution_max_rounds: 10
   tool_execution_max_steps: 6
   tool_execution_chunk_size: 2
@@ -231,6 +232,7 @@ Packaged prompts live under `src/outline_agent/assets/prompts/`, while user over
 | Field | Required | Default | Purpose | Notes |
 | --- | --- | --- | --- | --- |
 | `runtime.dry_run` | No | `false` | Simulates side effects without mutating Outline state. | Useful for testing plans and prompts safely. |
+| `runtime.collection_memory_document_title` | No | `MEMORY` | Title of the remote Outline document used for collection-wide memory. | The runtime lazily pulls this document on the first involved chat after startup, mirrors it into local `memory/MEMORY.md`, and only creates it on the first actual memory write. |
 | `runtime.tool_execution_max_rounds` | No | `10` | Maximum number of planning/replanning rounds for one request. | Increase this for more complex tool-driven tasks. |
 | `runtime.tool_execution_max_steps` | No | `6` | Maximum number of steps the planner may propose in one round. | This is the planner budget, not the total global task budget. |
 | `runtime.tool_execution_chunk_size` | No | `2` | Number of planned steps actually executed before replanning. | Smaller chunks improve adaptability after partial failures. |

docs/workspace-memory-and-prompts.md

@@ -9,6 +9,9 @@ This document describes the current runtime storage model after the document-mem
 - The only durable semantic memory layers are:
   - **collection memory**
   - **document memory**
+- Collection memory is **remote-first**:
+  - the Outline document titled `MEMORY` (configurable) is the user-facing source of truth
+  - `memory/MEMORY.md` is the local cache used for prompting and local inspection
 - Thread storage now keeps:
   - transcript
   - runtime state
@@ -22,6 +25,7 @@ For one collection, the workspace now looks like this:
 ```text
 <workspace_root>/<collection_id>/
   memory/
+    .collection.json
     00_SYSTEM.md
     MEMORY.md
     index.json
@@ -64,8 +68,9 @@ Purpose:
 Notes:
 
 - `00_SYSTEM.md` is included into the final system prompt as collection-local instructions
-- `MEMORY.md` stores collection facts, decisions, and working notes
+- `MEMORY.md` is the local cache of the remote collection memory document
 - `workspace/` is where tools read/write files and where generated artifacts live before upload
+- the remote Outline document title defaults to `MEMORY` and can be changed with `runtime.collection_memory_document_title`
 
 ### 2. Document layer
 
@@ -116,7 +121,9 @@ It is rebuilt from Outline comments on webhook events so deletes and edits are r
 
 ### Semantic memory truth
 
-- collection-wide memory truth: `memory/MEMORY.md`
+- collection-wide memory truth:
+  - remote source of truth: Outline document titled `MEMORY` by default
+  - local cache: `memory/MEMORY.md`
 - document-wide memory truth: `documents/<document_id>/MEMORY.md`
 
 These are model-maintained summaries, not raw transcript mirrors.
@@ -144,7 +151,17 @@ On a new comment:
 4. document comments are fetched
 5. thread transcript is rebuilt/synced into `comments.json`
 6. thread runtime state is refreshed from transcript
-7. if trigger conditions pass, the agent is invoked
+7. if trigger conditions pass, the runtime performs a one-time per-process pull/check for the collection memory document
+8. the agent is invoked
+
+### `documents.create` / `documents.update`
+
+If the affected document matches the configured collection memory document title:
+
+1. only collections already involved by the agent in the current process are considered
+2. the runtime fetches the remote document body
+3. the local cache `memory/MEMORY.md` is created or replaced
+4. `memory/index.json` is regenerated
 
 ### `comments.update`
 
@@ -199,6 +216,7 @@ Managed by:
 
 - `src/outline_agent/managers/memory_manager.py`
 - `src/outline_agent/managers/memory_action_manager.py`
+- `src/outline_agent/managers/collection_memory_sync.py`
 
 Two paths exist:
 
@@ -208,8 +226,9 @@ If the user explicitly asks to remember / forget / correct collection memory:
 
 1. router enables memory action path
 2. memory action manager proposes `add|update|delete|move`
-3. `memory/MEMORY.md` is edited directly
-4. `memory/index.json` is regenerated
+3. the remote collection memory document is created on demand or updated
+4. the local cache `memory/MEMORY.md` is refreshed
+5. `memory/index.json` is regenerated
 
 This path is user-controlled and intentional.
 
@@ -219,8 +238,9 @@ After a successful reply:
 
 1. the collection memory updater reviews the latest interaction
 2. it proposes durable collection-level entries
-3. `memory/MEMORY.md` is appended/updated
-4. `memory/index.json` is regenerated
+3. the remote collection memory document is created on demand or updated
+4. the local cache `memory/MEMORY.md` is refreshed
+5. `memory/index.json` is regenerated
 
 This should only retain reusable cross-document/cross-thread memory.
 
@@ -328,7 +348,7 @@ These prompts still have important hard-coded protocol/skeleton pieces in code.
 | `src/outline_agent/planning/tool_planner.py::UNIFIED_TOOL_PLANNER_SYSTEM_PROMPT` | protocol prompt | tool planner protocol/schema; policy addendum now comes from the internal prompt policy file / override path |
 | `src/outline_agent/planning/tool_planner.py::_build_user_prompt` | user prompt template | provides tool catalog, work-dir inventory, prior rounds, document memory, attachments, and comment context |
 | `src/outline_agent/state/workspace.py::INITIAL_SYSTEM_TEMPLATE` | initialization template | bootstrap template for collection `memory/00_SYSTEM.md` |
-| `src/outline_agent/state/workspace.py::INITIAL_MEMORY_TEMPLATE` | initialization template | bootstrap template for collection `memory/MEMORY.md` |
+| `src/outline_agent/state/workspace.py::INITIAL_MEMORY_TEMPLATE` | initialization template | template used when collection memory is first created locally/remotely |
 | `src/outline_agent/state/workspace.py::INITIAL_DOCUMENT_MEMORY_TEMPLATE` | initialization template | bootstrap template for document `documents/<document_id>/MEMORY.md` |
 
 ### What is not currently externalized

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "outline-agent"
-version = "0.2.2"
+version = "0.2.3"
 description = "Webhook-driven Outline thread agent"
 readme = "README.md"
 requires-python = ">=3.11"

src/outline_agent/__init__.py

@@ -2,4 +2,4 @@
 
 __all__ = ["__version__"]
 
-__version__ = "0.2.2"
+__version__ = "0.2.3"

src/outline_agent/app.py

@@ -71,6 +71,21 @@ async def lifespan(_: FastAPI):
         )
     elif settings.tool_use_enabled:
         logger.warning("ask_web_search is unavailable: {}", web_search_probe.reason or "unknown reason")
+    outline_client = None
+    store = None
+    processor = None
+    try:
+        outline_client, store = build_request_resources(settings)
+        processor = build_comment_processor(
+            settings=settings,
+            outline_client=outline_client,
+            store=store,
+        )
+    except Exception:
+        logger.exception("Failed to initialize shared Outline processor resources")
+    app.state.outline_client = outline_client
+    app.state.store = store
+    app.state.processor = processor
     yield
 
 
@@ -189,6 +204,8 @@ async def outline_webhook(request: Request) -> JSONResponse:
         "comments.create",
         "comments.update",
         "comments.delete",
+        "documents.create",
+        "documents.update",
         "documents.delete",
         "collections.delete",
     }:
@@ -208,15 +225,17 @@ async def outline_webhook(request: Request) -> JSONResponse:
     except ValidationError as exc:
         raise HTTPException(status_code=400, detail=f"invalid webhook payload: {exc}") from exc
 
-    outline_client = None
-    store = None
     try:
-        outline_client, store = build_request_resources(settings)
-        processor = build_comment_processor(
-            settings=settings,
-            outline_client=outline_client,
-            store=store,
-        )
+        outline_client = getattr(app.state, "outline_client", None)
+        store = getattr(app.state, "store", None)
+        processor = getattr(app.state, "processor", None)
+        if outline_client is None or store is None or processor is None:
+            outline_client, store = build_request_resources(settings)
+            processor = build_comment_processor(
+                settings=settings,
+                outline_client=outline_client,
+                store=store,
+            )
         result = await processor.handle(envelope)
         logger.info(
             "Webhook processed: action={}, reason={}, comment_id={}, document_id={}, collection_id={}",

src/outline_agent/clients/model_client.py

@@ -545,6 +545,9 @@ async def _iter_sse_events(response: httpx.Response) -> AsyncIterator[tuple[str
         if line.startswith(":"):
             continue
         if line.startswith("event:"):
+            if data_lines:
+                yield current_event, "\n".join(data_lines)
+                data_lines = []
             current_event = line[6:].strip() or None
             continue
         if line.startswith("data:"):

src/outline_agent/core/config.py

@@ -493,6 +493,7 @@ def settings_customise_sources(
     thread_turn_max_chars: int = 280
     memory_update_max_entries: int = 3
     memory_update_entry_max_chars: int = 180
+    collection_memory_document_title: str = "MEMORY"
     document_memory_update_enabled: bool = True
     document_memory_model_ref: str | None = None
     document_memory_summary_max_chars: int = 320