✦ Engram

Your photos, given a memory.

A self-hosted "second brain" for your life's photos — built on open-source Cognee.

Add a photo → it's understood, woven into a living knowledge graph, and connected to every related moment you've ever captured. Then ask your memories anything.

Self-hosted Cognee · AWS Bedrock · Knowledge Graph + Vectors · Next.js + FastAPI

---

Live demo: https://tolerance-maternity-bolt-friendly.trycloudflare.com

Blog: https://medium.com/@poojabhavani19/we-gave-our-photos-a-memory-ff90b354bb82

Video link: https://youtu.be/UnO_IYURbKs?si=K_Rdaxnc-4nDSMwa

Engram is self-hosted by design — the live link tunnels to a machine running the full stack (Cognee from source, Postgres+pgvector, local Kokoro voice). If the tunnel is offline, everything runs locally in two commands — see Run it locally — and the demo video shows the full experience.

---

Why Engram

Your camera roll is the most detailed diary you'll ever keep — and the one you never read. Engram turns that pile of flat, unsearchable images into a connected memory you can reason with: it understands what's in each photo, links people / places / moods / moments into a knowledge graph, and lets you ask, explore, and even teach it.

Who it's for. Engram is built around its most meaningful use: a private reminiscence companion — for an aging parent, someone whose memory is fading, or a family preserving a life story. It doesn't just show photos; it gently leads a person from one memory to a connected one they'd forgotten, read aloud, hands-free. Reminiscence is a clinically-grounded practice for memory care, and the most personal data of all should never leave the home — which is exactly why self-hosted, open-source Cognee is the heart of this project, not a footnote.

Most "AI memory" is a vector database doing nearest-neighbour search. Engram is different because Cognee gives it a real knowledge graph alongside vectors — vectors find what's similar, the graph understands what's connected. And it's 100% self-hosted: your memories never leave your machine.

Cognee's four-operation memory lifecycle — all live

Op	In Engram	How
remember()	Upload a photo → it joins the graph	cognee.add(...) + cognee.cognify(...) (background)
recall()	Ask questions that connect memories	cognee.search(GRAPH_COMPLETION), answer + the exact photos used
improve()	👍/👎 an answer → the memory gets smarter	cognee.session.add_feedback() + cognee.improve(session_ids=…)
forget()	Delete a memory → its subgraph leaves	cognee.forget(...)

Features

• Reminiscence Companion (the heart of Engram) — a calm, voice-led, hands-free session that walks the knowledge graph: it starts at one memory and follows the connections (a shared person, place or feeling) to lead someone gently to a related memory they'd forgotten. Built for memory care. (/reminisce)
• Forgotten Connection — graph serendipity: two memories that quietly rhyme through shared concepts, surfaced because only a graph could notice.
• Living knowledge graph — every photo's people, places, moods and objects, extracted by Cognee and visualised interactively (/graph).
• Concept Constellation — a physics-driven map of the concepts in your life; click a concept and its memories light up.
• Memory Connections — for any photo, traverse the graph to surface other memories that share concepts.
• Ask your memories — graph-grounded Q&A with multiple reasoning modes (graph / chain-of-thought / summary) and a self-improving feedback loop.
• City Lights — an ambient animated view of the places your memories wandered.
• Human-voice narration (ElevenLabs, Polly fallback).

Architecture

Photos ─▶ AWS Bedrock (Claude vision + Titan embeddings)
       ─▶ Cognee  ── cognify ──▶ Knowledge Graph (Kuzu)  +  Vectors (LanceDB)
       ─▶ FastAPI  ── remember / recall / improve / forget / graph / concepts / connections
       ─▶ Next.js  ── the Engram experience

• Memory: self-hosted Cognee on one Postgres (+pgvector) instance serving all three stores — relational, graph and vectors (COGNEE_STORE=postgres); flips to fully-local Kuzu + LanceDB + SQLite with COGNEE_STORE=local. No cloud memory service either way.
• LLM + embeddings: AWS Bedrock (Claude 3.5 Sonnet + Titan) via Cognee's LiteLLM layer — runs on existing credits, no per-call SaaS bill.
• Photo metadata: SQLite (canonical record for the UI); Cognee holds the semantic memory + graph.

Run it locally

Cognee runs self-hosted from source. Point COGNEE_SRC at a clone of topoteretes/cognee.

# Backend (FastAPI + self-hosted Cognee)  →  http://localhost:8000
cd api
./run.sh                     # PYTHONPATH=$COGNEE_SRC uvicorn app.main:app --port 8000

# (optional) seed sample memories through the API
python -m app.seed

# Frontend (Next.js 14)  →  http://localhost:3000
cd ../frontend
npm install
NEXT_PUBLIC_API_URL=http://localhost:8000 npm run dev

Config lives in api/app/main.py:configure_cognee() (brain paths + Bedrock providers, applied via Cognee's programmatic API) and api/.env (see comments). AWS credentials come from your ~/.aws profile.

The build story

I wrote up the experience — what Cognee is, how it helped, and the gotchas — in blog/.

---

_{Built for the WeMakeDevs × Cognee hackathon · Best Use of Open Source · Engram runs entirely on self-hosted, open-source Cognee.}