Cognee vs MemoryRouter: Feedback or Fact Corrections?

John Rood··3 min read

An agent gives the wrong answer. You click thumbs down. What should it remember?

The sources were useful, but the answer was too long. Or the answer was clear, but the office moved last month. A rating alone does not say which problem occurred.

Cognee has an answer-rating workflow and a separate factual-conflict review. MemoryRouter's MCP instructions tell the agent to save corrected information with enough context to identify what it replaces. The operation you need depends on what was wrong with the answer.

Cognee attaches a rating to a particular answer

Cognee's feedback guide starts with recall() and a session_id, with caching enabled. Get the session history, identify the answer's qa_id, and attach feedback with cognee.session.add_feedback. Scores run from 1 to 5, with an optional written explanation.

Recording the rating is not the graph update. Run improve() with the relevant session_ids to apply feedback weights to the graph elements used during retrieval. The guide also documents a background path through remember() with self_improvement=True.

Ranking has another control. feedback_influence uses DEFAULT_FEEDBACK_INFLUENCE, which is 0.0 by default. Learned weights do not change ranking until that influence is enabled globally or overridden for a recall.

For an application collecting answer-quality judgments, that gives you an interaction-linked signal and control over its effect on retrieval. Cognee's personalized-memory documentation separately describes per-user, per-dataset preferences, disabled by default. Those personal preferences are distinct from the shared feedback weights.

Cognee also handles factual disagreement. Its improve operation reviews conflicting entity facts by default, using their source documents and dates. It marks facts as current, superseded, or conflicting, and search exposes those marks. That review needs a configured LLM; it is skipped without one.

MemoryRouter saves the content of the correction

MemoryRouter's documented MCP interface exposes search_memories for recall and store_memory for storage. It does not list a tool for rating a Q&A record.

Its standing instructions explicitly include corrections among the details to save, with what the correction replaces when relevant. The useful payload is the changed information.

Consider a fictional correction:

The design review is now on Thursday. That replaces the Tuesday schedule we discussed before. The change starts next week.

Saving that statement gives a later session something specific to retrieve. Saving only “bad answer” leaves the next agent guessing whether the schedule was wrong or the response was poorly written.

The correction adds a memory. This documented workflow does not overwrite the older record or reduce its retrieval weight. An agent still has to retrieve the relevant context and interpret the correction. That is not automatic conflict resolution.

The MemoryRouter MCP setup guide explains how a compatible client connects to a vault. A generic MCP connection leaves tool invocation to the model. Connecting the server does not, by itself, prove that a tool ran on every message.

Ask for the kind of repair you need

If you want users to grade answers and have those grades influence retrieval, evaluate Cognee's feedback workflow. Check that improve() applied the feedback and that ranking influence is enabled. For graph-level factual adjudication, also inspect its conflict-review results.

If the problem is a personal agent repeating an outdated detail in another session, MemoryRouter's correction workflow fits that need. Give the correction its actual replacement value rather than a satisfaction score.

During evaluation, separate the two cases. Reject an answer for its presentation while keeping its facts correct. Then correct a fact while accepting the answer's presentation. Inspect the stored feedback or correction and the next retrieval.

For the factual-correction case, create a MemoryRouter vault.