developer-tools

Cognee Review: Graph-Backed Memory for AI Agents (2026)

Name: Cognee
Availability: InStock
Author: AI Demos

Inspectable graph-backed memory for AI agents, with strong provenance tracing but cautious update/delete behavior.

Visit Cognee

Graph-backed recallScope isolationUpdate cautionForget failure

TL;DR — our verdictUpdated July 2026 · 19 test artifacts

Strong observability and reuse, but lifecycle control still needs caution.

Where it wins

you need user, client, and project context to persist across sessions
you need graph views with visible provenance, IDs, and memory structure
you need separate account or project brains for multi-tenant agent workflows

Main limitation

you need immediate write-to-read consistency

Pricing (verified plans)

Free forever FreeWorkspace add-on $5/month per workspace

Strongest test artifacts

Persistent cross-session memory →Inspectable memory graph and provenance tracing →Scoped client and project memory isolation →

Our take

Cognee is a strong fit when you need inspectable, graph-backed memory for AI agents: the research used the hosted CloudClient path, and the tool preserved working-style, project, and client context across sessions while exposing enough provenance to debug what was recalled. The main caution is lifecycle control: an explicit ACME update did not fully retire stale SSO context, a conversational forget did not remove the stored preference, and Input 3 Session 4 still needs a direct screenshot/evidence-panel re-check before it can be finalized.

| Input | Session status | |---|---| | 1 | S1 memory created; S2 style reused; S3 boundary control | | 2 | S1 ACME context stored; S2 deeper troubleshooting; S3 update acknowledged; S4 stale SSO resurfaced; S5 BetaCorp isolated | | 3 | S1-S3 acknowledged; S4 needs verification; S5 isolated | | 4 | S1 memory created; S2 forget acknowledged; S3 inconclusive alone; forced probe fail |

Screen-recorded walkthrough used during the Cognee evaluation.

In-Depth Review

Our detailed analysis of Cognee — features, performance, and real-world testing.

AI Demos Team

Expert Reviewer

Verified Review

Feature-by-Feature Breakdown

Persistent cross-session memory

Strong

▾

Test Summary

Feature tested: Persistent cross-session memory

Result: Passed — Strong

Feature tested: Persistent cross-session memory

Result: Passed

Verdict: Strong

Expected behavior: Cognee stores durable context so later tasks can preserve the right tone and working assumptions. In testing, it held onto a concise, proof-first working style, kept a formal partner email from inheriting the internal-update voice, and accepted project context, rules, and direction updates inside the AI Demos.

Test case: Text prompt → Image

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Image): The assistant stored the working-style memory and confirmed the concise, proof-first, safe-next-step preference in the memory panel. — Cognee_Input1_Session1_Working_Style_Memory_Creation.png

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Image): The assistant stored the working-style memory and confirmed the concise, proof-first, safe-next-step preference in the memory panel. — Cognee_Input1_Session1_Working_Style_Memory_Creation.png

What changed: Text prompt transformed into Image

Test case: Text prompt → Image

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Image): The later internal update stayed short and research-oriented, and the recall panel linked back to the stored working-style memory rather than treating the session as new. — Cognee_Input1_Session2_Internal_Update_Graph_Recall.png

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Image): The later internal update stayed short and research-oriented, and the recall panel linked back to the stored working-style memory rather than treating the session as new. — Cognee_Input1_Session2_Internal_Update_Graph_Recall.png

What changed: Text prompt transformed into Image

Test case: Text prompt → Image

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Image): The formal partner email stayed professional and did not inherit the internal-update tone, which shows the memory was not over-applied to an unrelated task. — Cognee_Input1_Session3_Formal_Email_Boundary_Control.png

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Image): The formal partner email stayed professional and did not inherit the internal-update tone, which shows the memory was not over-applied to an unrelated task. — Cognee_Input1_Session3_Formal_Email_Boundary_Control.png

What changed: Text prompt transformed into Image

Test case: Text prompt → Image

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Image): Cognee accepted the project context and returned a short acknowledgment; the memory was captured, but the visible reply did not restate it in detail. — Cognee_Input3_Session1_Project_Context_Generic_Acknowledgment.png

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Image): Cognee accepted the project context and returned a short acknowledgment; the memory was captured, but the visible reply did not restate it in detail. — Cognee_Input3_Session1_Project_Context_Generic_Acknowledgment.png

What changed: Text prompt transformed into Image

Test case: Text prompt → Image

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Image): The rules were acknowledged, but the visible response remained brief rather than surfacing the full retrieved rule set. — Cognee_Input3_Session2_Project_Rules_Generic_Acknowledgment.png

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Image): The rules were acknowledged, but the visible response remained brief rather than surfacing the full retrieved rule set. — Cognee_Input3_Session2_Project_Rules_Generic_Acknowledgment.png

What changed: Text prompt transformed into Image

Test case: Text prompt → Image

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Image): Cognee acknowledged the direction change toward real workflows, showing that project direction updates were being retained across sessions. — Cognee_Input3_Session3_Direction_Change_Generic_Acknowledgment.png

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Image): Cognee acknowledged the direction change toward real workflows, showing that project direction updates were being retained across sessions. — Cognee_Input3_Session3_Direction_Change_Generic_Acknowledgment.png

What changed: Text prompt transformed into Image

Why it matters / Conclusion: Strong for normal persistence and boundary control, but the project-container branch surfaced mostly short acknowledgments rather than richly visible recall.

Cognee stores durable context so later tasks can preserve the right tone and working assumptions. In testing, it held onto a concise, proof-first working style, kept a formal partner email from inheriting the internal-update voice, and accepted project context, rules, and direction updates inside the AI Demos.

INPUT

Session 1, founder_001: define a working-style memory that keeps outputs short, direct, copy-paste ready, avoids overly polished writing, asks for proof before strong claims, and suggests the safest next step when a task is risky or unclear.

↓→

OUTPUT

The assistant stored the working-style memory and confirmed the concise, proof-first, safe-next-step preference in the memory panel.

INPUT

Session 2, founder_001: create a short internal update about testing AI memory tools and what to test next.

↓→

OUTPUT

The later internal update stayed short and research-oriented, and the recall panel linked back to the stored working-style memory rather than treating the session as new.

INPUT

Session 3, founder_001: write a formal email to a potential enterprise partner asking for a product demo next week.

↓→

OUTPUT

The formal partner email stayed professional and did not inherit the internal-update tone, which shows the memory was not over-applied to an unrelated task.

INPUT

Session 1, project_ai_demos: explain the AI Demos Memory for AI Agents project context and what it is trying to evaluate.

↓→

OUTPUT

Cognee accepted the project context and returned a short acknowledgment; the memory was captured, but the visible reply did not restate it in detail.

INPUT

Session 2, project_ai_demos: apply the project rules about proof, screenshots, relevance, scope control, and observability.

↓→

OUTPUT

The rules were acknowledged, but the visible response remained brief rather than surfacing the full retrieved rule set.

INPUT

Session 3, project_ai_demos: note the direction change toward real workflows like personal work, client relationships, and team handoffs.

↓→

OUTPUT

Cognee acknowledged the direction change toward real workflows, showing that project direction updates were being retained across sessions.

Bottom Line

Strong for normal persistence and boundary control, but the project-container branch surfaced mostly short acknowledgments rather than richly visible recall.

Inspectable memory graph and provenance tracing

Strong

▾

Test Summary

Feature tested: Inspectable memory graph and provenance tracing

Result: Passed — Strong

Feature tested: Inspectable memory graph and provenance tracing

Result: Passed

Verdict: Strong

Expected behavior: Cognee turns stored text into an inspectable graph and exposes Memory Schema and Mindmap views, including documents, chunks, entities, types, summaries, relation counts, source task names, and node provenance. The retrievals were auditable rather than opaque, and both views were available in testing.

Test case: Text prompt → Image

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Image): The schema view shows the extraction taxonomy and the layered structure Documents → Chunks → Entities → Types → Summaries, with automatic model, prompt, and ontology settings. — Memory Schema view.png

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Image): The schema view shows the extraction taxonomy and the layered structure Documents → Chunks → Entities → Types → Summaries, with automatic model, prompt, and ontology settings. — Memory Schema view.png

What changed: Text prompt transformed into Image

Test case: Text prompt → Image

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Image): The edge detail panel exposes source, target, relation type, ontology match status, and importance metadata, which is what makes the graph inspection auditable. — Mindmap overview edge detail.png

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Image): The edge detail panel exposes source, target, relation type, ontology match status, and importance metadata, which is what makes the graph inspection auditable. — Mindmap overview edge detail.png

What changed: Text prompt transformed into Image

Test case: Text prompt → Image

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Image): The node inspector reveals the chunk text plus provenance fields such as task, pipeline, user, relation counts, and ontology notes. — Mindmap node connections.png

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Image): The node inspector reveals the chunk text plus provenance fields such as task, pipeline, user, relation counts, and ontology notes. — Mindmap node connections.png

What changed: Text prompt transformed into Image

Why it matters / Conclusion: This is the standout strength of Cognee: it is unusually debuggable for a memory layer.

Cognee turns stored text into an inspectable graph and exposes Memory Schema and Mindmap views, including documents, chunks, entities, types, summaries, relation counts, source task names, and node provenance. The retrievals were auditable rather than opaque, and both views were available in testing.

INPUT

Open the Memory Schema view for the stored founder_001 memories after processing the sessions.

↓→

OUTPUT

The schema view shows the extraction taxonomy and the layered structure Documents → Chunks → Entities → Types → Summaries, with automatic model, prompt, and ontology settings.

INPUT

Inspect the Mindmap overview edge details for one of the stored graph relationships.

↓→

OUTPUT

The edge detail panel exposes source, target, relation type, ontology match status, and importance metadata, which is what makes the graph inspection auditable.

INPUT

Open a document chunk node in the Mindmap to inspect its connections and provenance.

↓→

OUTPUT

The node inspector reveals the chunk text plus provenance fields such as task, pipeline, user, relation counts, and ontology notes.

Bottom Line

This is the standout strength of Cognee: it is unusually debuggable for a memory layer.

Scoped client and project memory isolation

Strong

▾

Test Summary

Feature tested: Scoped client and project memory isolation

Result: Passed — Strong

Feature tested: Scoped client and project memory isolation

Result: Passed

Verdict: Strong

Expected behavior: Cognee keeps client and project containers separate so one customer's history does not bleed into another customer's reply and unrelated projects stay isolated. In testing, ACME support continuity moved forward, BetaCorp got a fresh first-contact reply, and a separate project stayed partitioned.

Test case: Text prompt → Image

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Image): The ACME client context was stored under its own account container, with the session panel showing the ACME-specific thread and recalled context. — Cognee_Input2_Session1_ACME_Context_Graph_Storage.png

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Image): The ACME client context was stored under its own account container, with the session panel showing the ACME-specific thread and recalled context. — Cognee_Input2_Session1_ACME_Context_Graph_Storage.png

What changed: Text prompt transformed into Image

Test case: Text prompt → Image

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Image): Cognee moved beyond the already-tried steps and suggested deeper SSO diagnostics such as IdP metadata, ACS URL, entity ID, and debug logging. — Cognee_Input2_Session2_ACME_Deeper_SSO_Troubleshooting.png

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Image): Cognee moved beyond the already-tried steps and suggested deeper SSO diagnostics such as IdP metadata, ACS URL, entity ID, and debug logging. — Cognee_Input2_Session2_ACME_Deeper_SSO_Troubleshooting.png

What changed: Text prompt transformed into Image

Test case: Text prompt → Image

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Image): BetaCorp received a fresh first-contact support reply with no ACME history bleeding into the new account. — Cognee_Input2_Session5_BetaCorp_Scope_Isolation.png

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Image): BetaCorp received a fresh first-contact support reply with no ACME history bleeding into the new account. — Cognee_Input2_Session5_BetaCorp_Scope_Isolation.png

What changed: Text prompt transformed into Image

Test case: Text prompt → Image

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Image): The unrelated sales project stayed on its own topic and did not inherit the AI Demos rules or the BetaCorp login scenario. — Cognee_Input3_Session5_Unrelated_Project_Scope_Isolation.png

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Image): The unrelated sales project stayed on its own topic and did not inherit the AI Demos rules or the BetaCorp login scenario. — Cognee_Input3_Session5_Unrelated_Project_Scope_Isolation.png

What changed: Text prompt transformed into Image

Why it matters / Conclusion: Scope separation looked reliable in the tested containers, and the ACME support thread advanced without leaking into BetaCorp or the unrelated project.

Cognee keeps client and project containers separate so one customer's history does not bleed into another customer's reply and unrelated projects stay isolated. In testing, ACME support continuity moved forward, BetaCorp got a fresh first-contact reply, and a separate project stayed partitioned.

INPUT

Session 1, client_acme_001: ACME is a client using the support assistant; they prefer clear next steps and already tried password reset, clearing browser cache, and switching browsers.

↓→

OUTPUT

The ACME client context was stored under its own account container, with the session panel showing the ACME-specific thread and recalled context.

INPUT

Session 2, client_acme_001: ACME still cannot log in with SSO; draft the next support reply and move to the next useful step.

↓→

OUTPUT

Cognee moved beyond the already-tried steps and suggested deeper SSO diagnostics such as IdP metadata, ACS URL, entity ID, and debug logging.

INPUT

Session 5, client_beta_002: BetaCorp is a new client and their users cannot log in for the first time; draft the first support reply.

↓→

OUTPUT

BetaCorp received a fresh first-contact support reply with no ACME history bleeding into the new account.

INPUT

Session 5, unrelated_sales_project: create a short kickoff note for a different sales email agent project.

↓→

OUTPUT

The unrelated sales project stayed on its own topic and did not inherit the AI Demos rules or the BetaCorp login scenario.

Bottom Line

Scope separation looked reliable in the tested containers, and the ACME support thread advanced without leaking into BetaCorp or the unrelated project.

Memory update and correction handling

Weak

▾

Test Summary

Feature tested: Memory update and correction handling

Result: Partial — Weak

Cognee can acknowledge corrections and accept updated context, but the tests show that an explicit update may not fully retire the old context. In testing, an ACME switch away from SSO was acknowledged, yet a later login reply still drifted back toward the older troubleshooting path.

INPUT

Session 3, client_acme_001: update the client memory so ACME is no longer using SSO for this rollout and should not be treated as an active SSO issue unless they mention it again.

↓→

OUTPUT

Cognee acknowledged the update and said ACME would no longer be treated as an active SSO case.

INPUT

Session 4, client_acme_001: ACME says some users are still having trouble logging in during launch testing; draft the next support reply.

↓→

OUTPUT

Despite the update, the follow-up reply still moved into SSO-related troubleshooting, showing stale context resurfacing.

INPUT

Session 4, client_acme_001: same launch-testing follow-up reply, with the conversation and memory panel visible together.

↓→

OUTPUT

This paired screenshot makes the stale SSO resurfacing easier to audit because the reply and memory panel are both visible.

Bottom Line

Acknowledgment works, but conflict resolution is weak because older context can still dominate later replies.

Delete / forget memory control

Failure

▾

Test Summary

Feature tested: Delete / forget memory control

Result: Failed — Failure

Cognee accepts a conversational forget request, but the stored preference remained retrievable under a forced probe. In testing, the chat reply changed while the underlying memory did not disappear from the graph layer.

INPUT

Session 1, delete_memory_001: remember that future updates should use a very formal corporate tone.

↓→

OUTPUT

The formal-tone preference was created and confirmed as stored memory.

INPUT

Session 2, delete_memory_001: forget the formal corporate tone preference and do not keep using it.

↓→

OUTPUT

Cognee acknowledged the forget request conversationally, but this screen alone does not prove deletion.

INPUT

Forced retrieval probe: what writing style preference do you remember for me?

↓→

OUTPUT

The forced probe still returned the original formal corporate tone preference, proving the memory remained retrievable.

Bottom Line

Confirmed failure: the forget request was acknowledged in chat, but the underlying memory was still available for retrieval.

Free forever plan plus usage-based workspace pricing

All testing in this report was done on the free plan.

TESTED

Free forever

Free

1 workspace, about $2.50 in included credits (1,000,000 tokens), no card required.

Workspace add-on

$5/month per workspace

Unlimited users and unlimited API calls beyond the first free workspace; usage billed at $2.50 per 1,000,000 tokens processed.

All testing in this report was done on the free plan; no paid workspace was purchased.

✓ Use This If

●you need user, client, and project context to persist across sessions

●you need graph views with visible provenance, IDs, and memory structure

●you need separate account or project brains for multi-tenant agent workflows

●you can tolerate a short background processing window before new memory is retrievable

✕ Skip This If

●you need immediate write-to-read consistency

●you need update requests to reliably retire stale context without extra checks

●you need a conversational forget request to guarantee actual deletion

developer-toolsagent-platformstext

Yes. The tests showed it preserving working-style preferences, project direction, and client support context across separate sessions.

Yes, through two different surfaces. This report's custom test interface displayed recall responses with evidence chunks, document IDs, and dataset IDs — data returned by Cognee's own API, not a Cognee product screen. Separately, Cognee's own native dashboard provides the Memory Schema and Mindmap views for inspecting stored graph structure directly.

Yes. BetaCorp received a fresh first-contact reply without ACME's SSO history bleeding into it.

No. The update was acknowledged, but the next ACME login reply still resurfaced SSO-oriented troubleshooting.

Not by itself in this test. The forget request was acknowledged, but a forced retrieval probe still returned the original formal-tone preference.

The report states a free forever plan with 1 workspace and about $2.50 in included credits (1,000,000 tokens), no card required. It also states a $5/month per workspace add-on with usage billed at $2.50 per 1,000,000 tokens processed.

Through the official Python SDK (`cognee`) using `CloudClient` against a hosted Cognee Cloud tenant with a tenant URL and API key. The report also says REST API and MCP server access were available.

It is still the main open question. The handoff note is structurally complete, but the testing-direction and artifact details appear to drift toward ACME/BetaCorp login content, so it should be re-checked against the screenshot and evidence panel before being finalized.