Cognee review: choosing graph memory and verifying cross-session recall
A documentation-based analysis of permanent memory, session storage and local models, with a pilot plan covering updates, deletion and access isolation.
AgentIndex · Published 2026-11-04 · Updated 2026-11-04
Cognee is worth evaluating when agents repeatedly have to reconstruct relationships between documents, earlier decisions and code. It organizes information into retrievable graph memory. Adoption should depend on whether those relationships help recover correct evidence and whether the team can manage changes and deletion.
For a handful of stable project conventions, a maintained instruction file is often easier to inspect. Consider a memory platform when growing sources, relationships and users make manual context management burdensome, then compare the benefit with the additional operational work.
Scope and version
On November 4, 2026, we checked the project README, PyPI metadata, the remember, recall and forget documentation, and the current remember implementation. PyPI listed version 1.6.6 with Python 3.10–3.14 support. Package information
This is a documentation- and source-based selection review. We did not install models, build a business knowledge graph, measure retrieval accuracy or execute the example below. The synthetic documents and acceptance table describe a proposed evaluation, rather than results we obtained.
Session memory and permanent memory need different tests
Without a session identifier, remember enters the permanent-memory ingestion path. With one, it first stores content in the session cache; a subsequent improvement process can bridge that content into the permanent graph, and that behavior can be disabled. Remember reference
This distinction changes what counts as successful memory. Reading information immediately after writing it in the same session demonstrates that one retrieval path works. To assess durable memory, start another session, confirm the permanent write or bridge has completed, and inspect whether the original evidence is still retrievable. Record the write destination, processing status and read path rather than treating a successful function return as the entire acceptance test.
Automatic improvement should also be observable. A conversational statement should not silently replace an authoritative business document merely because an agent encountered it most recently.
Local processing and generated answers are separate decisions
The README describes a no-LLM-key text path using a local GLiNER extractor and local embeddings, with model downloads on first use. Local operation still requires storage, compute and preparation. Generated answers and media processing that needs vision or transcription require further configuration. Local-mode overview
For internal information, record the provider separately for extraction, embeddings, retrieval and answer generation. A local model at one stage does not establish that every other stage is local. Start with a small set of nonsensitive text before introducing business material.
Inspect evidence before judging answers
Recall can use session-related paths or retrieve from the permanent graph with dataset scoping. Its documentation describes source tags on results and distinguishes retrieval without a usable LLM from generated answers. Recall reference
Evaluate these as two questions: did retrieval return the right evidence, and did the answer accurately use it? If the evidence is wrong, a more articulate model may conceal the problem. If evidence is right but the answer is wrong, investigate generation instructions, citations and abstention before rebuilding the graph.
A useful pilot: changing a project’s release rules
Create three short synthetic documents. An old deployment note says releases happen on Fridays. A later decision changes the November schedule to Tuesdays. An incident review describes an exception during the end-of-month freeze. Attach a date, source and scope to each record.
Ask about the current release day, the previous month’s rule, the freeze exception and the evidence supporting each answer. These tasks expose whether the workflow distinguishes obsolete information, finds exceptions and admits missing evidence. They are proposed test cases, not the release policy of a real organization.
| Check | Proposed procedure | What to observe |
|---|---|---|
| Cross-session retrieval | Finish ingestion, stop the calling process and query in a new session | Whether evidence remains available and which source supplies it |
| Updates | Add a dated decision while retaining the older record | Whether current and historical rules remain distinguishable |
| Deletion | Delete a document from a test dataset and repeat relevant queries | Whether original content or derived results still appear |
| Isolation | Give two users different test material and permissions | Whether unauthorized evidence can be retrieved |
| Unknowns | Ask about a rule absent from every source | Whether the system acknowledges the lack of support |
Forget distinguishes document deletion, dataset deletion and resetting graph/vector memory while retaining raw files for reprocessing. A memory-only reset therefore must not be mistaken for complete removal of source material. Forget reference
A minimal starting example
This adapts the README’s interface to a separate dataset with synthetic text. We have not run it, asserted a return schema or measured its accuracy:
python -m venv .venv
source .venv/bin/activate
python -m pip install "cognee[gliner]==1.6.6"
import asyncio
import cognee
async def main():
await cognee.remember(
"Example project: the November release day is Tuesday.",
dataset_name="release_review_demo",
)
results = await cognee.recall(
"Which day is the November release?",
datasets=["release_review_demo"],
)
print(results)
asyncio.run(main())
Check installation and model preparation first, ingestion completion second, and returned evidence third. Then expand to multiple documents, process restarts and new sessions. Measure first-time downloads, initial graph construction and subsequent queries separately; a cold start should not represent every future request.
Compare with simpler approaches
| Need | A useful starting point | What Cognee must demonstrate |
|---|---|---|
| A few stable conventions | Maintained project instructions | Less context-maintenance work |
| Relevant passages from documents | An observable document-retrieval pipeline | Additional useful evidence from relationships |
| Connections across documents, code and past decisions | A graph-memory pilot | Reliable updates, provenance and cross-session retrieval |
Our selection criterion is a demonstrable benefit from relationships. If a simpler workflow already supplies correct evidence, retain it until the pilot shows a reason to change. If you adopt graph memory, plan updates, deletion, permission isolation and recovery as part of the same project, rather than stopping at one successful question and answer.
Related reading: Impeccable detector review, Parlant customer-agent selection, and choosing open-source RAG.