Ending AI Amnesia in Meetings: Building the Fireflies.ai Connector for Cognee

guthikonda naveen — Tue, 13 Oct 2026 11:25:16 +0000

Have you ever asked an AI assistant about a critical decision made during yesterday's team sync, only to be met with a generic apology or complete hallucination?

AI agents typically excel at reading documents from Notion or Google Drive and indexing code repositories on GitHub. But here is the reality of modern engineering and product teams: over 80% of high-impact technical, architectural, and business commitments happen verbally in meetings.

When your AI agents lack access to spoken conversations, they suffer from conversational amnesia. They cannot recall who committed to fixing a bug, what trade-offs were discussed for an upcoming release, or why an architecture decision was approved.

For Mergetober—the month-long open source celebration with Cognee and WeMakeDevs—I built and contributed the Fireflies.ai Data-Source Connector (Pull Request #375) to solve this problem.

In this post, I will walk you through how the connector bridges meeting speech and Cognee's knowledge graph, the architectural challenges we solved, and how you can use it in your own agentic workflows.

What is Cognee?

Cognee is the open-source memory layer for AI agents. Rather than dumping raw text into a standard vector database and relying solely on semantic similarity search (which loses entity relationships and structural context), Cognee blends:

Knowledge Graphs: Nodes representing entities, concepts, and people, linked by typed edges representing real-world relationships.
Vector Embeddings: Fine-grained semantic embeddings for flexible retrieval.
Graph Retrieval & Completion: Enabling agents to traverse multi-hop connections to answer complex reasoning queries.

The Vision: Turning Spoken Words into a Living Knowledge Graph

Integrating meeting intelligence is much more complex than fetching plain text. A raw transcript is noisy, messy, and lacks semantic structure. If an LLM reads "John: I'll handle the API rate limits", without context, the agent doesn't know which John, what his email is, or what exact task was committed.

Our Fireflies connector ingests transcripts, executive overviews, action items, and speaker-attributed turns from Fireflies.ai directly into Cognee.

Key Architectural Solutions

1. True Speaker Attribution

In Fireflies, speaker utterances often contain only display names or raw labels (Speaker 1). Our connector implements a speaker-resolution engine that correlates meeting turns against the verified meeting attendee list (Name (email)).

This allows Cognee's entity extraction pipeline to establish explicit, unambiguous graph edges:

[Alice (alice@example.com)] ──(committed to)──> [Action Item: Update Rate Limits]
[Bob (bob@example.com)] ──(spoke about)──> [Architecture Migration to Kubernetes]

2. Document-Mode Routing

Cognee supports different ingestion pipelines. For meeting notes, we declared DOCUMENT_SOURCE_ATTR = "fireflies". This directs synthesized meeting summaries and dialogues through Cognee's natural language entity extraction (GLiNER / LLM extraction) rather than rigid tabular database schemas.

3. Incremental Synchronization

Nobody wants to re-ingest hundreds of hours of historical meetings on every run. The connector uses a high-watermark timestamp cursor (fromDate) persisted across runs. On subsequent executions, only meetings recorded since the last sync are fetched.

4. Forget-on-Delete Reconciliations

What happens when a meeting is deleted from Fireflies for privacy or compliance? If your memory layer doesn't reconcile deletions, zombie nodes persist in your graph. Our connector emits _deleted: True tombstones on merge, which triggers Cognee's orphan_cleanup mechanism to automatically purge obsolete entities. We also included a safe-sweep guard to protect against transient empty API responses.

5. Plan-Quota & Rate-Limit Resilience

Fireflies enforces API quotas (such as 50 requests/day on free developer tiers). The connector handles HTTP 429 and rate-limiting headers with intelligent exponential backoff and Retry-After header tracking, preventing ingestion jobs from crashing.

Quickstart: Using the Fireflies Connector

Here is how simple it is to give your Cognee agents persistent meeting memory:

1. Installation

pip install cognee-community-connector-fireflies
# Or using uv:
uv add cognee-community-connector-fireflies

2. Ingest & Recall in Python

import asyncio
import os
import cognee
from cognee_community_connector_fireflies import fireflies_source

DATASET_NAME = "team_meeting_memory"

async def main():
    # 1. Initialize the Fireflies source (uses FIREFLIES_API_KEY)
    source = fireflies_source(
        days_back=14,               # Ingest meetings from the last 2 weeks
        include_action_items=True,  # Extract action items
        include_sentences=True,     # Extract full speaker dialogue
    )

    # 2. Ingest into Cognee Knowledge Graph
    print("🧠 Ingesting meeting transcripts into Cognee memory...")
    await cognee.remember(
        source,
        dataset_name=DATASET_NAME,
        primary_key="id",
        write_disposition="merge",
    )

    # 3. Ask your AI agent about meeting commitments
    query = "Who committed to updating the rate-limit documentation in our sync?"
    result = await cognee.search(
        query_text=query,
        query_type=cognee.SearchType.GRAPH_COMPLETION,
        datasets=[DATASET_NAME],
    )
    print("\n🔍 Recall Result:\n", result)

if __name__ == "__main__":
    asyncio.run(main())

100% Mocked Offline Test Suite

Open-source contributions must be reliable and maintainable. To ensure contributors and CI environments can test without needing proprietary Fireflies credentials, the entire test suite in tests/test_fireflies.py is 100% offline and mocked:

Comprehensive test coverage (16/16 tests passing in 0.19s).
Full coverage for speaker resolution, pagination, tombstone emissions, cursor persistence, and rate-limit backoff.
Full compliance with the Topoteretes Developer Certificate of Origin (DCO).

Links & Acknowledgements

Pull Request: topoteretes/cognee-community #375
GitHub Repository: topoteretes/cognee
Mergetober Hackathon: WeMakeDevs Mergetober

Huge thanks to the Cognee team and WeMakeDevs for organizing Mergetober and building such an innovative ecosystem for AI memory. Let's make AI agents remember what actually matters!

DEV Community: guthikonda naveen