Cognee boundary

Type: agentic-system-analyses/types/agentic-system-boundary.md

Boundary and evidence

This run analyses Cognee's shipped memory platform to explain how content becomes retained knowledge, how sessions and feedback maintain it, and how callers retrieve and reuse it. The target is a memory/knowledge/context-engineering system, not a claim that Cognee owns every consuming agent's execution loop. The public package exposes ingestion, knowledge processing, retrieval, memory maintenance and session APIs (SRC-1, cognee/__init__.py).

This module exposes the main functions and objects intended for direct use through import cognee. It groups the stable V1 API, memory-oriented V2 API, visualization helpers, tracing utilities, migration helpers, and session models behind a single package-level entrypoint. --- cognee/__init__.py

Cognee qualifies even though some routes need no generative LLM. The frozen documentation distinguishes local extraction and retrieval from generated answers, and implementation includes model-provider dispatch and model-facing maintenance instructions (SRC-1, README.md, cognee/infrastructure/llm/LLMGateway.py, cognee/infrastructure/llm/prompts/session_distillation_curator_system.txt).

Text ingestion, retrieval, and session storage work without an LLM. LLM-dependent improvement stages skip automatically. --- README.md

You are reading one slice of a finished working session and proposing durable lessons to persist into a knowledge graph. The slice is in chronological order and contains: --- cognee/infrastructure/llm/prompts/session_distillation_curator_system.txt

The functional boundary includes the Python library; its V1 and memory-oriented APIs; ingestion, extraction, code indexing, storage, authorization, retrieval and answer construction; sessions, feedback, distillation, improvement, deletion and migration; shipped prompts and maintenance instructions; and shipped CLI, HTTP, MCP, UI, starter-kit and example callers. Shipped evaluators and tests are also inside the source boundary as evidence of evaluation design and intended contracts. Packaging exclusions do not make repository-shipped callers external. Newly inspected material at this same commit remains SRC-1 evidence.

Initial tracing establishes permanent ingestion through add and cognify, session ingestion with an optional improvement bridge, retrieval from session and graph state, and stage-based maintenance (SRC-1, cognee/api/v1/remember/remember.py, cognee/api/v1/recall/recall.py, cognee/api/v1/improve/improve.py). The memory decorator retrieves context before invoking the wrapped function and persists a trace afterward (SRC-1, cognee/modules/agent_memory/decorator.py). These are source-level responsibilities, not observed successful operation.

context.memory_context = await retrieve_memory_context(context) result = await fn(args, *kwargs) --- cognee/modules/agent_memory/decorator.py

The shipped Company Brain driver belongs to the target's memory-use evidence: it sets storage roots, ingests text and code, distills a session and offers a separate invocation that queries saved memory. It is not excluded merely because it calls the library (SRC-1, examples/cookbooks/company_brain/docs_code_conversations/company_brain_demo.py). Its separate recall branch is explicit:

if args.recall_only: await recall_saved_memory(cognee) return --- examples/cookbooks/company_brain/docs_code_conversations/company_brain_demo.py

The MCP client likewise belongs to the shipped interface boundary. Its recall implementation has both an HTTP route and a local Python call; its improve implementation forwards maintenance requests in either mode (SRC-1, cognee-mcp/src/cognee_client.py). The CLI's declared entry point and remember/recall command implementations were also located (SRC-1, pyproject.toml, cognee/cli/commands/remember_command.py, cognee/cli/commands/recall_command.py).

Coverage and remaining loop limits

The boundary kind is complete artifact, partial loop. The complete repository artifact is pinned, but initial inspection does not establish whole-system coverage of every shipped state-producing and state-consuming path. In particular, frontend, connector, backend-worker and evaluator paths remain available but not substantively traced. This is an inspection limit, not an access gap. The run must not infer whole-system closure, uniform interface behavior or evaluation effectiveness from the central API trace alone.

All rows describe SRC-1; repository paths are initially inspected evidence locations or explicit remaining inspection work, not additional source identities. No shipped material responsibility is excluded from later inspection at the pinned commit.

External model providers, model weights and their download services, database engine internals, Enola internals, separately hosted Cognee Cloud, and agent plugins or applications outside this repository are excluded implementations. The shipped adapters, configuration and interface calls remain included. Dependency declarations establish required integration surfaces, not external implementations (SRC-1, pyproject.toml); README links to a separate integrations repository do not freeze that repository (SRC-1, README.md). These exclusions prevent conclusions about provider accuracy, external service behavior, independently distributed plugin capture policies, and whether an arbitrary host actually uses recalled material to change its actions. The wrapped application function is not attributed to Cognee merely because the decorator invokes it.

The frozen source is the acquired Git object, unchanged: https://github.com/topoteretes/cognee at f154b32a9e544dac73e7ddf2010327cf9988edcd. Checkout HEAD was checked against that pin. The applicability cutoff is 2026-10-14, the supplied run date. Only this revision supplies evidence; live documentation, linked repositories, releases and later changes are not frozen inputs. The overall tier is code-grounded. No target, example or test was executed and no provider was contacted. Initial evidence supports implementation and doctrine/design, not observed-run or causal-experiment claims.

Source register