cognee 1.2.1.dev20260622__tar.gz → 1.2.2.dev0__tar.gz

{cognee-1.2.1.dev20260622 → cognee-1.2.2.dev0}/.env.template

@@ -467,13 +467,6 @@ WEB_SCRAPER_MAX_DELAY=10.0
 #VECTOR_DB_PASSWORD=""
 #VECTOR_DB_SUBPROCESS_ENABLED=true
 
-# -- Observability (Langfuse) -------------------------------------------------
-# Set MONITORING_TOOL=langfuse and provide keys to enable LLM tracing.
-#MONITORING_TOOL="langfuse"
-#LANGFUSE_PUBLIC_KEY=""
-#LANGFUSE_SECRET_KEY=""
-#LANGFUSE_HOST=""
-
 # -- AWS / Bedrock extras (in addition to the AWS section above) --------------
 #AWS_PROFILE_NAME=""
 #AWS_BEDROCK_RUNTIME_ENDPOINT=""
@@ -491,6 +484,52 @@ WEB_SCRAPER_MAX_DELAY=10.0
 #FASTAPI_USERS_RESET_PASSWORD_TOKEN_SECRET="change_me_in_production"
 
 
+################################################################################
+#  Docker / MCP Runtime
+#  Configure the cognee API image (cognee/cognee) and the MCP image
+#  (cognee/cognee-mcp) when running `docker run` / `docker compose`.
+#  Unless noted "read by the app", these are consumed by the container
+#  entrypoints/compose and have defaults baked into the images — set them only
+#  to override. (docker-compose.yml already sets sensible values for most.)
+################################################################################
+
+# -- API server (cognee/cognee image) ----------------------------------------
+# CORS allow-list for the FastAPI server: comma-separated origins. Read by the
+# app (cognee/api/client.py). Default '*' (all origins) — set explicit domains
+# in production.
+#CORS_ALLOWED_ORIGINS="https://yourdomain.com,https://another.com"
+# Server bind/port inside the container (entrypoint defaults shown).
+#HTTP_PORT=8000
+#BIND_ADDRESS=0.0.0.0
+
+# -- MCP server (cognee/cognee-mcp image) -------------------------------------
+# Transport the MCP container serves. The Docker image reads TRANSPORT_MODE;
+# the direct `cognee-mcp` CLI uses --transport instead.
+#TRANSPORT_MODE=stdio          # stdio | sse | http
+# Comma-separated optional extras to pip-install at container startup.
+#EXTRAS=aws,postgres
+# MCP "API mode": point the MCP server at an already-running cognee API server.
+#API_URL=http://localhost:8000
+#API_TOKEN=""
+# MCP "Cloud mode": point the MCP server at a managed Cognee Cloud instance.
+# These are the canonical cloud-connection variables, shared across serve(),
+# push(), the MCP server, and sync. COGNEE_CLOUD_API_URL / COGNEE_CLOUD_AUTH_TOKEN
+# (above) remain as deprecated fallbacks.
+#COGNEE_SERVICE_URL=""
+#COGNEE_API_KEY=""
+
+# -- Debug (both images) ------------------------------------------------------
+# DEBUG=true together with ENV in {dev,local} starts the container under
+# debugpy, listening on DEBUG_PORT. ENV is the canonical environment variable
+# (set it in the Dev/Debug section above); ENVIRONMENT is a deprecated alias
+# still accepted by the container entrypoints.
+#DEBUG=false
+#DEBUG_PORT=5678
+
+# -- Frontend (cognee-frontend image / compose `ui` profile) ------------------
+#NEXT_PUBLIC_BACKEND_API_URL=http://localhost:8000
+
+
 ###############################################################################
 # TIER 4 — EXAMPLE PROVIDER OVERRIDES (commented out)
 # Uncomment + fill values to switch providers.

{cognee-1.2.1.dev20260622 → cognee-1.2.2.dev0}/CLAUDE.md

@@ -54,7 +54,7 @@ pre-commit install
 - **evals** - Evaluation tools
 - **deepeval** - DeepEval testing framework
 - **posthog** - PostHog analytics
-- **monitoring** - Sentry + Langfuse observability
+- **tracing** - OpenTelemetry tracing
 - **distributed** - Modal distributed execution
 - **dev** - All development tools (pytest, ty, ruff, etc.)
 - **debug** - Debugpy for debugging
@@ -101,7 +101,7 @@ ty check .
 ### Running Cognee
 ```bash
 # Using Python SDK
-python examples/python/simple_example.py
+uv run python examples/demos/simple_cognee_example.py
 
 # Using CLI
 cognee-cli add "Your text here"

{cognee-1.2.1.dev20260622 → cognee-1.2.2.dev0}/CONTRIBUTING.md

@@ -112,27 +112,7 @@ Copy `.env.template` to `.env` and provide your OPENAI_API_KEY as LLM_API_KEY
 Make sure to run ```shell uv sync ``` in the root cloned folder or set up a virtual environment to run cognee
 
 ```shell
-python examples/python/simple_example.py
-```
-or
-
-```shell
-uv run python examples/python/simple_example.py
-```
-
-### Running Simple Example
-
-Copy `.env.template` to `.env` and provide your OPENAI_API_KEY as LLM_API_KEY
-
-Make sure to run ```shell uv sync ``` in the root cloned folder or set up a virtual environment to run cognee
-
-```shell
-python cognee/cognee/examples/python/simple_example.py
-```
-or
-
-```shell
-uv run python cognee/cognee/examples/python/simple_example.py
+uv run python examples/demos/simple_cognee_example.py
 ```
 
 ## 4. 📤 Submitting Changes

{cognee-1.2.1.dev20260622 → cognee-1.2.2.dev0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: cognee
-Version: 1.2.1.dev20260622
+Version: 1.2.2.dev0
 Summary: Cognee - is a library for enriching LLM context with a semantic layer for better understanding and reasoning.
 Project-URL: Homepage, https://www.cognee.ai
 Project-URL: Repository, https://github.com/topoteretes/cognee
@@ -135,13 +135,6 @@ Requires-Dist: llama-index-core<0.15,>=0.14.20; extra == 'llama-index'
 Provides-Extra: mistral
 Requires-Dist: mistral-common<2,>=1.5.2; extra == 'mistral'
 Requires-Dist: mistralai<2,>=1.9.10; extra == 'mistral'
-Provides-Extra: monitoring
-Requires-Dist: langfuse<3,>=2.32.0; extra == 'monitoring'
-Requires-Dist: opentelemetry-api<2,>=1.20.0; extra == 'monitoring'
-Requires-Dist: opentelemetry-exporter-otlp-proto-grpc<2,>=1.20.0; extra == 'monitoring'
-Requires-Dist: opentelemetry-exporter-otlp-proto-http<2,>=1.20.0; extra == 'monitoring'
-Requires-Dist: opentelemetry-sdk<2,>=1.20.0; extra == 'monitoring'
-Requires-Dist: sentry-sdk[fastapi]<3,>=2.9.0; extra == 'monitoring'
 Provides-Extra: neo4j
 Requires-Dist: neo4j<6,>=5.28.0; extra == 'neo4j'
 Provides-Extra: neptune

{cognee-1.2.1.dev20260622 → cognee-1.2.2.dev0}/cognee/api/client.py

@@ -57,21 +57,6 @@ from cognee.modules.users.methods.get_authenticated_user import REQUIRE_AUTHENTI
 setup_logging()
 logger = get_logger()
 
-if os.getenv("ENV", "prod") == "prod":
-    try:
-        import sentry_sdk
-
-        sentry_sdk.init(
-            dsn=os.getenv("SENTRY_REPORTING_URL"),
-            traces_sample_rate=1.0,
-            profiles_sample_rate=1.0,
-        )
-    except ImportError:
-        logger.info(
-            "Sentry SDK not available. Install with 'pip install cognee\"[monitoring]\"' to enable error monitoring."
-        )
-
-
 app_environment = os.getenv("ENV", "prod")
 
 

{cognee-1.2.1.dev20260622 → cognee-1.2.2.dev0}/cognee/api/v1/add/add.py

@@ -263,7 +263,7 @@ async def add(
         pipeline_name="add_pipeline",
         vector_db_config=vector_db_config,
         graph_db_config=graph_db_config,
-        use_pipeline_cache=True,
+        use_pipeline_cache=False,
         incremental_loading=incremental_loading,
         data_per_batch=data_per_batch,
         llm_config=llm_config,

{cognee-1.2.1.dev20260622 → cognee-1.2.2.dev0}/cognee/api/v1/cognify/cognify.py

@@ -270,7 +270,7 @@ async def cognify(
             vector_db_config=vector_db_config,
             graph_db_config=graph_db_config,
             incremental_loading=incremental_loading,
-            use_pipeline_cache=True,
+            use_pipeline_cache=False,
             pipeline_name="cognify_pipeline",
             data_per_batch=data_per_batch,
             rollback_handler=cognify_rollback_handler,

{cognee-1.2.1.dev20260622 → cognee-1.2.2.dev0}/cognee/api/v1/improve/improve.py

@@ -155,12 +155,10 @@ async def improve(
                 )
                 return {}
             acquired_lock_for = sole_session
-        else:
-            release_improve_lock = None  # type: ignore[assignment]
 
-        # Stage 1 & 2: bridge sessions into the permanent graph
-        if session_ids:
-            try:
+        try:
+            # Stage 1 & 2: bridge sessions into the permanent graph
+            if session_ids:
                 await _bridge_sessions(
                     dataset=dataset,
                     session_ids=session_ids,
@@ -182,6 +180,12 @@ async def improve(
                 )
                 stages_run.append("persist_trace_steps")
 
+                # Stage 2b2: distill each session's agent traces into agent-profile
+                # session-context lessons (the LLM batch pass) before distillation, so
+                # those lessons are available as gated guidance for stage 2c.
+                if await _extract_agent_context(session_ids=session_ids, user=user):
+                    stages_run.append("extract_agent_context")
+
                 # Stage 2c: distill each session's gated guidance into curated,
                 # entity-anchored lessons and add+cognify them into the graph.
                 # This is what lets remember(session, self_improvement=True)
@@ -194,62 +198,56 @@ async def improve(
                 )
                 if distilled:
                     stages_run.append("distill_sessions")
-            except Exception:
-                if acquired_lock_for:
-                    from cognee.infrastructure.locks import release_improve_lock
-
-                    await release_improve_lock(acquired_lock_for)
-                raise
-
-        # Stage 3: default enrichment (triplet embeddings)
-        from cognee.modules.memify import memify
 
-        if "node_type" not in kwargs or kwargs.get("node_type") is None:
-            from cognee.modules.engine.models.node_set import NodeSet
+            # Stage 3: default enrichment (triplet embeddings)
+            from cognee.modules.memify import memify
 
-            kwargs["node_type"] = NodeSet
-
-        result = await memify(
-            dataset=dataset,
-            node_name=node_name,
-            user=user,
-            run_in_background=run_in_background,
-            **kwargs,
-        )
-        stages_run.append("memify_enrichment")
+            if "node_type" not in kwargs or kwargs.get("node_type") is None:
+                from cognee.modules.engine.models.node_set import NodeSet
 
-        if build_global_context_index:
-            if run_in_background:
-                logger.warning(
-                    "improve: global context index skipped in background mode "
-                    "because ordered background pipeline chaining is not supported"
-                )
-            else:
-                global_context_index_updated = await _build_global_context_index(
-                    dataset=dataset,
-                    user=user,
-                )
-                if global_context_index_updated:
-                    stages_run.append("global_context_index")
+                kwargs["node_type"] = NodeSet
 
-        # Stage 5: sync enriched graph back to session cache (incremental)
-        # Skip when running in background — stage 3 hasn't completed yet
-        if session_ids and not run_in_background:
-            await _sync_graph_to_sessions(
+            result = await memify(
                 dataset=dataset,
-                session_ids=session_ids,
+                node_name=node_name,
                 user=user,
+                run_in_background=run_in_background,
+                **kwargs,
             )
-            stages_run.append("sync_graph_to_sessions")
-
-        span.set_attribute(COGNEE_IMPROVE_STAGES, ",".join(stages_run))
+            stages_run.append("memify_enrichment")
+
+            if build_global_context_index:
+                if run_in_background:
+                    logger.warning(
+                        "improve: global context index skipped in background mode "
+                        "because ordered background pipeline chaining is not supported"
+                    )
+                else:
+                    global_context_index_updated = await _build_global_context_index(
+                        dataset=dataset,
+                        user=user,
+                    )
+                    if global_context_index_updated:
+                        stages_run.append("global_context_index")
+
+            # Stage 5: sync enriched graph back to session cache (incremental)
+            # Skip when running in background — stage 3 hasn't completed yet
+            if session_ids and not run_in_background:
+                await _sync_graph_to_sessions(
+                    dataset=dataset,
+                    session_ids=session_ids,
+                    user=user,
+                )
+                stages_run.append("sync_graph_to_sessions")
 
-        if acquired_lock_for:
-            from cognee.infrastructure.locks import release_improve_lock
+            span.set_attribute(COGNEE_IMPROVE_STAGES, ",".join(stages_run))
 
-            await release_improve_lock(acquired_lock_for)
+            return result
+        finally:
+            if acquired_lock_for:
+                from cognee.infrastructure.locks import release_improve_lock
 
-        return result
+                await release_improve_lock(acquired_lock_for)
 
 
 async def _build_global_context_index(
@@ -337,6 +335,47 @@ async def _bridge_sessions(
         logger.warning("improve: session persistence failed (non-fatal): %s", e)
 
 
+async def _extract_agent_context(
+    session_ids: List[str],
+    user,
+) -> int:
+    """Flush pending trace windows into agent-profile lessons before distillation.
+
+    Delegates to ``agent_context_extraction.extract_pending_agent_context`` per session, which
+    shares the same watermark used by mid-session trace extraction. ``min_new_traces=1`` makes
+    improve/session-end flush any remaining unprocessed traces before distillation. Gated on
+    automatic session context and best-effort/fail-open: an error on one session never blocks the
+    others or the rest of ``improve()``. Returns the number of lessons created/linked.
+    """
+    from cognee.infrastructure.session.agent_context_extraction import (
+        extract_pending_agent_context,
+    )
+    from cognee.infrastructure.session.get_session_manager import get_session_manager
+
+    session_manager = get_session_manager()
+    if not session_manager.is_available or not session_manager.is_auto_feedback_enabled():
+        return 0
+
+    user_id = str(user.id)
+    touched = 0
+    for session_id in session_ids:
+        try:
+            ids = await extract_pending_agent_context(
+                session_manager=session_manager,
+                user_id=user_id,
+                session_id=session_id,
+                min_new_traces=1,
+            )
+            touched += len(ids)
+        except Exception as e:
+            logger.warning(
+                "improve: agent-context extraction failed for '%s' (non-fatal): %s",
+                session_id,
+                e,
+            )
+    return touched
+
+
 async def _distill_sessions(
     dataset: Union[str, UUID],
     session_ids: List[str],

{cognee-1.2.1.dev20260622 → cognee-1.2.2.dev0}/cognee/api/v1/recall/recall.py

@@ -30,6 +30,7 @@ from cognee.modules.recall.types.RecallResponse import (
     ResponseGraphContextEntry,
     ResponseGraphEntry,
     ResponseQAEntry,
+    ResponseSessionContextEntry,
 )
 from cognee.modules.recall.types.SearchResultItem import SearchResultItem
 from cognee.modules.search.models.SearchResultPayload import SearchResultPayload
@@ -313,6 +314,49 @@ async def _fetch_graph_context(
     return [ResponseGraphContextEntry(content=snapshot, source="graph_context")]
 
 
+async def _fetch_session_context(
+    query_text: str,
+    session_id: str,
+    context_profile: str,
+    user: str | None = None,
+) -> list[ResponseSessionContextEntry]:
+    """Render active session-context lessons for one profile as a one-item list.
+
+    Read-only: uses the deterministic builder with ``stamp_served=False`` so it never updates
+    served metadata. Returns an empty list when no lessons match the profile.
+    """
+    from cognee.infrastructure.session.get_session_manager import get_session_manager
+    from cognee.infrastructure.session.session_context_builder import build_active_context_block
+
+    caller_user_id = await _resolve_user_id(user)
+    if not caller_user_id:
+        return []
+    cache_user_id = await _resolve_session_cache_user_id(session_id, caller_user_id)
+    if not cache_user_id:
+        return []
+
+    sm = get_session_manager()
+    if not sm.is_available:
+        return []
+
+    block, _served = await build_active_context_block(
+        session_manager=sm,
+        user_id=cache_user_id,
+        session_id=session_id,
+        query=query_text,
+        context_profile=context_profile,
+        stamp_served=False,
+    )
+    if not block:
+        return []
+
+    return [
+        ResponseSessionContextEntry(
+            content=block, context_profile=context_profile, source="session_context"
+        )
+    ]
+
+
 async def recall(
     query_text: str,
     query_type: SearchType | None = None,
@@ -328,6 +372,7 @@ async def recall(
     node_name_filter_operator: str = "OR",
     only_context: bool = False,
     session_id: str | None = None,
+    context_profile: str = "qa",
     wide_search_top_k: int | None = 100,
     triplet_distance_penalty: float | None = 6.5,
     feedback_influence: float = 0.0,
@@ -442,6 +487,7 @@ async def recall(
                 node_name=node_name,
                 only_context=only_context,
                 session_id=session_id,
+                context_profile=context_profile,
                 verbose=verbose,
                 include_references=include_references,
             )
@@ -480,6 +526,18 @@ async def recall(
                 return []
             return list(await _fetch_graph_context(session_id=session_id, user=user))
 
+        async def _run_session_context() -> list[RecallResponse]:
+            if not session_id:
+                return []
+            return list(
+                await _fetch_session_context(
+                    query_text=query_text,
+                    session_id=session_id,
+                    context_profile=context_profile,
+                    user=user,
+                )
+            )
+
         async def _run_graph() -> list[RecallResponse]:
             nonlocal user, dataset_ids
 
@@ -570,6 +628,7 @@ async def recall(
             "session": _run_session,
             "trace": _run_trace,
             "graph_context": _run_graph_context,
+            "session_context": _run_session_context,
             "graph": _run_graph,
         }
 

{cognee-1.2.1.dev20260622 → cognee-1.2.2.dev0}/cognee/api/v1/recall/routers/get_recall_router.py

@@ -82,8 +82,15 @@ class RecallPayloadDTO(InDTO):
         examples=[None],
         description=(
             "Which memory sources to include: 'graph', 'session', 'trace', "
-            "'graph_context', 'all', 'auto', or a list of these. Defaults to "
-            "'auto' (session first when session_id is set, else graph)."
+            "'graph_context', 'session_context', 'all', 'auto', or a list of these. Defaults "
+            "to 'auto' (session first when session_id is set, else graph)."
+        ),
+    )
+    context_profile: str = Field(
+        default="qa",
+        description=(
+            "Profile to render for the 'session_context' scope: 'qa' (conversational) or "
+            "'agent' (tool/workflow). Ignored by other scopes."
         ),
     )
 
@@ -181,6 +188,7 @@ def get_recall_router() -> APIRouter:
                 only_context=payload.only_context,
                 session_id=payload.session_id,
                 scope=payload.scope,
+                context_profile=payload.context_profile,
                 include_references=payload.include_references,
             )
             return jsonable_encoder(results)

{cognee-1.2.1.dev20260622 → cognee-1.2.2.dev0}/cognee/api/v1/serve/cloud_client.py

@@ -194,6 +194,8 @@ class CloudClient:
             payload["session_id"] = kwargs["session_id"]
         if kwargs.get("scope") is not None:
             payload["scope"] = kwargs["scope"]
+        if kwargs.get("context_profile") is not None:
+            payload["context_profile"] = kwargs["context_profile"]
         if kwargs.get("include_references") is not None:
             payload["include_references"] = kwargs["include_references"]
 

{cognee-1.2.1.dev20260622 → cognee-1.2.2.dev0}/cognee/api/v1/sync/sync.py

@@ -564,13 +564,25 @@ async def _get_file_size(file_path: str) -> int:
 
 
 async def _get_cloud_base_url() -> str:
-    """Get Cognee Cloud API base URL."""
-    return os.getenv("COGNEE_CLOUD_API_URL", "http://localhost:8001")
+    """Get Cognee Cloud API base URL.
+
+    Canonical: COGNEE_SERVICE_URL (shared with serve()/push()/MCP).
+    COGNEE_CLOUD_API_URL is accepted as a deprecated fallback.
+    """
+    return (
+        os.getenv("COGNEE_SERVICE_URL")
+        or os.getenv("COGNEE_CLOUD_API_URL")
+        or "http://localhost:8001"
+    )
 
 
 async def _get_cloud_auth_token(user: User) -> str:
-    """Get authentication token for Cognee Cloud API."""
-    return os.getenv("COGNEE_CLOUD_AUTH_TOKEN", "your-auth-token")
+    """Get authentication token for Cognee Cloud API.
+
+    Canonical: COGNEE_API_KEY (shared with serve()/push()/MCP).
+    COGNEE_CLOUD_AUTH_TOKEN is accepted as a deprecated fallback.
+    """
+    return os.getenv("COGNEE_API_KEY") or os.getenv("COGNEE_CLOUD_AUTH_TOKEN") or "your-auth-token"
 
 
 async def _check_hashes_diff(

{cognee-1.2.1.dev20260622 → cognee-1.2.2.dev0}/cognee/base_config.py

@@ -34,15 +34,8 @@ class BaseConfig(BaseSettings):
         self.system_root_directory = ensure_absolute_path(self.system_root_directory)
         self.logs_root_directory = ensure_absolute_path(self.logs_root_directory)
 
-        # Set monitoring tool based on available keys
-        if self.langfuse_public_key and self.langfuse_secret_key:
-            self.monitoring_tool = Observer.LANGFUSE
-
         return self
 
-    langfuse_public_key: Optional[str] = os.getenv("LANGFUSE_PUBLIC_KEY")
-    langfuse_secret_key: Optional[str] = os.getenv("LANGFUSE_SECRET_KEY")
-    langfuse_host: Optional[str] = os.getenv("LANGFUSE_HOST")
     default_user_email: Optional[str] = os.getenv("DEFAULT_USER_EMAIL")
     default_user_password: Optional[str] = os.getenv("DEFAULT_USER_PASSWORD")
 

{cognee-1.2.1.dev20260622 → cognee-1.2.2.dev0}/cognee/cli/user_resolution.py

@@ -19,7 +19,7 @@ async def resolve_cli_user(user_id: Optional[str] = None, strict: bool = False):
     from cognee.modules.users.methods import get_default_user
 
     if not user_id:
-        return await get_default_user()
+        return await _get_default_user_with_recovery()
 
     try:
         uid = UUID(user_id)
@@ -44,6 +44,22 @@ async def resolve_cli_user(user_id: Optional[str] = None, strict: bool = False):
             f"User {uid} not found — falling back to the default user.  "
             f"The --user-id will not provide isolation."
         )
+        return await _get_default_user_with_recovery()
+
+
+async def _get_default_user_with_recovery():
+    """Try get_default_user(); on DatabaseNotCreatedError run migrations and retry."""
+    from cognee.modules.users.methods import get_default_user
+    from cognee.infrastructure.databases.exceptions import DatabaseNotCreatedError
+
+    try:
+        return await get_default_user()
+    except DatabaseNotCreatedError:
+        from cognee.infrastructure.databases.relational import get_relational_engine
+        from cognee.modules.migrations.startup import run_migrations
+
+        await run_migrations()
+
         return await get_default_user()
 
 

{cognee-1.2.1.dev20260622 → cognee-1.2.2.dev0}/cognee/infrastructure/llm/config.py

@@ -208,14 +208,13 @@ class LLMConfig(BaseSettings):
             "EMBEDDING_PROVIDER": is_env_set("EMBEDDING_PROVIDER"),
             "EMBEDDING_MODEL": is_env_set("EMBEDDING_MODEL"),
             "EMBEDDING_DIMENSIONS": is_env_set("EMBEDDING_DIMENSIONS"),
-            "HUGGINGFACE_TOKENIZER": is_env_set("HUGGINGFACE_TOKENIZER"),
         }
         if any(embedding_env_vars.values()) and not all(embedding_env_vars.values()):
             missing_embed = [key for key, is_set in embedding_env_vars.items() if not is_set]
             raise ValueError(
                 "You have set some but not all of the required environment variables "
                 "for embeddings (EMBEDDING_PROVIDER, EMBEDDING_MODEL, "
-                "EMBEDDING_DIMENSIONS, HUGGINGFACE_TOKENIZER). Missing: "
+                "EMBEDDING_DIMENSIONS). Missing: "
                 f"{missing_embed}"
             )
 

cognee-1.2.2.dev0/cognee/infrastructure/llm/prompts/agent_context_extraction_system.txt

@@ -0,0 +1,23 @@
+You read a batch of an agent's tool/action traces from one session and extract a small set of
+reusable lessons that would help the agent on future turns in this environment.
+
+The input may contain two sections:
+- EXISTING LESSONS: lessons already saved for this session. Do NOT repeat or reword these — only
+  propose lessons that add genuinely new information.
+- TRACES: the trace steps to learn from. Each line shows the tool/function, its status, a one-line
+  feedback summary, any error, and a truncated output.
+
+Return a list of lessons (at most 5). Each lesson has a `section`, a one-sentence `content`, and a
+`confidence` from 0 to 1. Only emit a lesson you are confident is durable and reusable; omit
+anything one-off, trivial, specific to a single value, or already covered by EXISTING LESSONS. It
+is fine to return an empty list.
+
+Choose `section` by what the lesson is:
+- tool_rules: a reusable constraint or safe/required way to invoke a tool.
+- workflow_state: where the task stands and what must survive context compaction to continue it.
+- success_patterns: an approach or sequence that solved the user's request and is worth repeating.
+- failure_lessons: an error, retry, or rejected path and what to avoid or do differently.
+- environment_facts: a discovered fact about the repository, runtime, configuration, or environment.
+
+Keep each `content` to one short, self-contained sentence — a takeaway, not a description of one
+specific call. Do not restate raw output. Prefer fewer, higher-quality lessons.

cognee-1.2.2.dev0/cognee/infrastructure/llm/retry_config.py

@@ -0,0 +1,24 @@
+"""Shared tenacity retry policy for LLM structured-output calls.
+
+Used by every structured-output framework — the litellm/instructor adapters and
+the BAML integration alike. Each ``acreate_structured_output`` retries until BOTH
+floors are satisfied: at least ``LLM_MIN_RETRY_ATTEMPTS`` attempts AND at least
+``LLM_MIN_RETRY_SECONDS`` of elapsed wall-clock time.
+
+``&`` builds tenacity's ``stop_all`` predicate, which stops only once *every*
+sub-condition holds (``|`` / ``stop_any`` would stop at whichever floor is hit
+first). The predicate is stateless — it reads everything off the per-call retry
+state — so this single instance is safe to share across every retry decorator.
+"""
+
+from tenacity import stop_after_attempt, stop_after_delay
+
+# Minimum number of attempts before the call is allowed to give up.
+LLM_MIN_RETRY_ATTEMPTS = 2
+# Minimum elapsed seconds before the call is allowed to give up.
+LLM_MIN_RETRY_SECONDS = 240
+
+# Stop retrying only once BOTH the attempt floor AND the time floor are met.
+llm_retry_stop_condition = stop_after_attempt(LLM_MIN_RETRY_ATTEMPTS) & stop_after_delay(
+    LLM_MIN_RETRY_SECONDS
+)

{cognee-1.2.1.dev20260622 → cognee-1.2.2.dev0}/cognee/infrastructure/llm/structured_output_framework/baml/baml_src/extraction/acreate_structured_output.py

@@ -7,11 +7,11 @@ from tenacity import (
     before_sleep_log,
     retry,
     retry_if_not_exception_type,
-    stop_after_delay,
     wait_exponential_jitter,
 )
 
 from cognee.infrastructure.llm.config import get_llm_config
+from cognee.infrastructure.llm.retry_config import llm_retry_stop_condition
 from cognee.infrastructure.llm.structured_output_framework.baml.baml_client import b
 from cognee.infrastructure.llm.structured_output_framework.baml.baml_client.type_builder import (
     TypeBuilder,
@@ -28,7 +28,7 @@ T = TypeVar("T", bound="BaseModel | str")
 
 
 @retry(
-    stop=stop_after_delay(128),
+    stop=llm_retry_stop_condition,
     wait=wait_exponential_jitter(8, 128),
     retry=retry_if_not_exception_type(asyncio.CancelledError),
     before_sleep=before_sleep_log(logger, logging.WARNING),