DEV Community

Zahid Hamdule
Zahid Hamdule

Posted on

Giving AI a Permanent Memory: How We Built Founder's Fuel Using Cognee

LLMs are brilliant, but they suffer from a fatal flaw: they forget.

Every time you open a new chat window, upload a new file, or start a new session, the model starts from scratch. Standard retrieval techniques (like naive Vector RAG) help retrieve documents, but they miss the structural context. They don't understand how your pitch deck links to customer tickets, how investor conversations relate to sector trends, or how relationships build over time.

To solve this, we built Founder’s Fuel—a unified dashboard for startup founders to evaluate ideas, match with venture capital, and track signals. More importantly, we built it with Cognee Cloud as the central, graph-aware cognitive memory layer, giving our AI a permanent, structured memory.

Here is how we did it.

The Vision: A Connected Startup Workspace

Founder’s Fuel solves fundraising and development friction across six key features. By integrating Cognee, we didn't just build separate tools; we built a shared knowledge graph where information from one tool enriches the others.

                  ┌──────────────────────────────┐
                  │   VC Co-Pilot Unified Chat   │
                  └──────────────┬───────────────┘
                                 │ (Queries all 5 datasets in parallel)
  ┌──────────────────────────────┼──────────────────────────────┐
  ▼                              ▼                              ▼
┌──────────────────┐   ┌──────────────────┐   ┌──────────────────┐
│ Pitch Deck Graph │   │  Investor Graph  │   │ Customer Signals │
│  (dataset: memory)│   │(dataset: match)  │   │ (dataset: signals)│
└──────────────────┘   └──────────────────┘   └──────────────────┘
  ▲                                                             ▲
  └─────────────── (Linked dynamically via Cognee) ─────────────┘
Enter fullscreen mode Exit fullscreen mode

Under the Hood: The 5 Cognee memory datasets

1. Naive RAG to Graph Evaluation (Startup Evaluator)

Traditional evaluators match startup ideas against rigid lists. In Founder's Fuel, we parse uploaded files, chunk the content, and ingest it into Cognee's pitch_deck_memory dataset. Cognee builds semantic nodes (e.g., GTMStrategy, CompetitiveMoat, TeamCredentials), mapping relationships dynamically to build an explainable scorecard.

# Ingesting unstructured slide text into Cognee
cognee.add(slide_text, dataset_name = "pitch_deck_memory")
cognee.cognify(["pitch_deck_memory"])
Enter fullscreen mode Exit fullscreen mode

2. Semantic Sector Traversal (Smart Investor Matcher)

Naively searching a CSV for investor matches fails to capture abstract overlaps. We mapped 2,200+ venture capital profiles into Cognee's investor_graph. By calling search with GRAPH_COMPLETION, the matcher traverses stage, sector, and ticket-size nodes, computing matching scores based on graph proximity.

3. Customer Signal Engine

Customer feedback is highly relational. A discord bug report links to a product segment which links to a feature ticket. We ingest Discord, Slack, and ticketing logs into the customer_signals dataset. Cognee links issues and feature requests, giving founders a clear priority graph.

4. Investor Relationship CRM Memory

Fundraising is a multi-month loop of calls, meetings, and emails. We built a custom glassmorphic CRM dashboard where founders log interactions. Cognee indexes the notes and sentiment verdicts into investor_crm_memory, forming a structured timeline of investor relations that can be queried naturally.

# Logging interaction note
payload = f"VC: Sequoia Capital\nDate: 2026-07-05\nVerdict: Positive\nNotes: Loved our evaluation graph."
cognee.add(payload, "investor_crm_memory")
cognee.cognify(["investor_crm_memory"])
Enter fullscreen mode Exit fullscreen mode

5. VC Co-Pilot Unified Chat

The crown jewel of our implementation. Instead of isolated chatbots, the unified chat page queries all five Cognee datasets in parallel using Python's ThreadPoolExecutor. The chatbot synthesizes the response, attributing its answers with custom badges representing the exact graph dataset sourced.

# Querying all 5 Cognee datasets concurrently
with ThreadPoolExecutor(max_workers=5) as executor:
    futures = {
        executor.submit(cognee_search, user_msg, ds_name): ds_name
        for ds_name in ["founders_fuel_kb", "investor_graph", "customer_signals", "pitch_deck_memory", "investor_crm_memory"]
    }
Enter fullscreen mode Exit fullscreen mode

Why Cognee is a Game-Changer for Hackathons

Building AI memory is usually a tedious process involving setting up Neo4j databases, vector indices, and custom indexing pipelines. Cognee made it incredibly simple. With a single, unified REST API and SDK, we were able to:

  1. Automatically construct graphs from raw text.
  2. Ingest and search multiple datasets concurrently.
  3. Perform explainable, hybrid search out of the box.

By using Cognee Cloud, we gave our application a brain that grows smarter as the founder uploads more pitch decks, receives more customer feedback, and logs more meetings.

Check out **Founder’s Fuel* on GitHub and experience the cognitive memory layer firsthand!*

Top comments (0)