diff --git a/CHANGELOG.md b/CHANGELOG.md
index a2c3705e..4a0f8eb2 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -9,6 +9,13 @@
   - Supports `all` argument to download/upload all datasets at once
   - Use `--force` flag to overwrite existing databases
   - Avoids lengthy database rebuild times for users running benchmarks
+- **Dynamic Session Context**: Compressed conversation history for multi-turn chat
+  - New `SessionContext` model stores summarized conversation state instead of raw Q&A history
+  - Background LLM-based summarization runs after each `ask` tool call (non-blocking)
+  - Previous summarization tasks are cancelled when new ones start
+  - Research graph receives compact context (~1,000-2,000 tokens) instead of raw qa_history (potentially thousands of tokens)
+  - New `session_context` field on `ChatSessionState` synced via AG-UI state protocol
+  - Chat TUI: New context modal (`Ctrl+O`) to view current session context
 
 ### Changed
 
@@ -28,7 +35,6 @@
 - **v0.25.0 Migration Failure**: Fixed "Table 'documents' already exists" error during migration caused by held table references preventing `drop_table()` from succeeding. Added recovery logic to restore documents from staging table if a previous migration attempt failed mid-way.
 
 ## [0.26.8] - 2026-01-22
-
 - **Jina Reranker v3**: Added support for Jina reranking with API mode (`provider: jina`) and local inference (`provider: jina-local`, requires `[jina]` extra)
 - **Model Downloads**: `download-models` now pre-downloads HuggingFace models for `sentence-transformers`, `mxbai`, and `jina-local`
 - **Reranker Factory**: Removed unreliable `id(config)`-based caching from `get_reranker()`; factory now always instantiates fresh
diff --git a/haiku_rag_slim/haiku/rag/agents/chat/agent.py b/haiku_rag_slim/haiku/rag/agents/chat/agent.py
index 113fc84a..1e1fcc01 100644
--- a/haiku_rag_slim/haiku/rag/agents/chat/agent.py
+++ b/haiku_rag_slim/haiku/rag/agents/chat/agent.py
@@ -25,8 +25,8 @@ from haiku.rag.utils import get_model
 
 logger = logging.getLogger(__name__)
 
-# Track background tasks to prevent garbage collection
-_background_tasks: set[asyncio.Task[None]] = set()
+# Track summarization tasks per session to allow cancellation
+_summarization_tasks: dict[str, asyncio.Task[None]] = {}
 
 
 async def _update_context_background(
@@ -42,6 +42,8 @@ async def _update_context_background(
             session_state=session_state,
         )
         logger.debug("Session context updated")
+    except asyncio.CancelledError:
+        logger.debug("Session context summarization cancelled")
     except Exception:
         logger.exception("Failed to update session context")
 
@@ -289,6 +291,11 @@ def create_chat_agent(config: AppConfig) -> Agent[ChatDeps, str]:
                 ]
 
             # Spawn background task to update session context
+            # Cancel any previous summarization for this session
+            session_id = ctx.deps.session_state.session_id
+            if session_id in _summarization_tasks:
+                _summarization_tasks[session_id].cancel()
+
             task = asyncio.create_task(
                 _update_context_background(
                     qa_history=list(ctx.deps.session_state.qa_history),
@@ -296,8 +303,8 @@ def create_chat_agent(config: AppConfig) -> Agent[ChatDeps, str]:
                     session_state=ctx.deps.session_state,
                 )
             )
-            _background_tasks.add(task)
-            task.add_done_callback(_background_tasks.discard)
+            _summarization_tasks[session_id] = task
+            task.add_done_callback(lambda t: _summarization_tasks.pop(session_id, None))
 
         # Build new state with citations AND accumulated qa_history
         new_state = ChatSessionState(