Memory that knows when to update itself.
Your AI forgets last night and confidently repeats stale facts. Engram is the self-reconciling memory layer that catches contradictions the moment they appear, and forgets only what has stopped mattering.
EngramResolve contradictions
Superseded decisions in your build
Everything you read, write, and build
ChatGPT export
Claude sessionArticlesYouTubeGitHubspecs.pdfChatGPT exportClaude sessionArticlesYouTubeGitHubspecs.pdfChatGPT exportClaude sessionArticlesYouTubeGitHubspecs.pdfChatGPT exportClaude sessionArticlesYouTube
GeminiRepositoriesPDF docsNotion notesTranscriptsWeb pagesGeminiRepositoriesPDF docsNotion notesTranscriptsWeb pagesGeminiRepositoriesPDF docsNotion notesTranscriptsWeb pagesGeminiRepositoriesPDF docsThe memory lifecycle
Four operations, total recall.
remember recall improve forget are all load-bearing, wired to both the local Cognee SDK and a hosted Cognee Cloud tenant, with automatic local fallback.
Ingest repos, PDFs, articles, chat exports and transcripts, structured into one graph.
Ask in plain language. Cognee routes between semantic search and deep graph traversal.
Re-enrich the graph, reinforce confirmed facts, and adapt weights from your feedback.
Decay unreinforced nodes over time and prune what no longer deserves to be remembered.
Ingest from anywhere
Everything you read, write, and build flows into one graph.
Repos, PDFs, articles, YouTube transcripts, and ChatGPT or Claude exports are all compiled by Cognee into a single reconciled memory.
ChatGPT ExportClaude Session
ChatGPT ExportClaude SessionGitHub Commitspecs.pdfNotion NotesYouTube AudioWeb ArticleSee it think
Ask your memory. Get a grounded answer.
A quick taste of recall. Type a question or pick one, and watch Engram answer from the graph, with the source it used.
Ask something below, or tap a suggestion. This is a live-typed demo of graph-grounded recall.
What makes it different
Most memory tools stop at recall. This one decides what still deserves trust.
Five source types, natively
Ingests PDFs, GitHub repositories, ChatGPT and Claude exports, YouTube transcripts, and web articles, with your chat turns remembered as you go.
Reconciliation engine
Validates every new belief against your graph in under two seconds using schema-level contradiction checks.
Time-aware decay
Confidence decays continuously and unreinforced nodes are pruned once they drop below the threshold.
A living 3D knowledge graph
A force-directed visualizer maps your memory as a weighted network. Nodes grow with connections; edges show supersedes and contradicts, with real-time provenance tracing.
Hover a node to trace its connections.
Temporal diffs
Ask what changed since March and get a diff of added nodes, superseded beliefs, and new decisions.
The Recap: where is my context?
A morning-after digest of your memory. For any window, Engram stitches every lifecycle operation into one grounded narrative.
Bring your own key
Connect Groq, OpenAI, or Gemini. Keys are validated live and encrypted at rest with Fernet.
Similarity RAG vs Engram
Recall is not enough. Memory has to stay honest.
Similarity RAG
- Retrieves stale facts with full confidence
- No contradiction detection at ingestion
- Static text embeddings, no structure
- Context grows unbounded and noisy
Engram
- Catches contradictions the moment they appear
- Confidence decays and stale facts are pruned
- Deterministic Cognee knowledge graph
- Forgets only what has stopped mattering
How it works
From raw context to a memory that maintains itself.
Ingest
Point Engram at a repo, PDF, article, video, or chat export. It extracts the meaning and structures it into the graph.
Reconcile
New evidence is checked against what you already know. Contradictions surface instantly for a quick decision.
Recall
Ask across every session. Answers are graph-grounded and time-aware, with a diff of what changed.
Decay
Unreinforced beliefs fade and get pruned, so recall stays fast, lean, and trustworthy.
Questions
Good to know
No. Similarity-based RAG relies on static embeddings and happily retrieves stale facts. Engram uses Cognee to compile context into a deterministic graph and runs schema checks at ingestion to catch factual contradictions before they enter long-term memory.
The older belief is deactivated and its confidence drops toward zero. Every reconciliation decision is logged, so you can run temporal diffs like what changed since last week at any time.
GitHub repositories, local PDFs, ChatGPT and Claude conversation exports, web articles, and YouTube transcripts, plus your own chat turns as you use the app.
Only to the LLM provider you configure. Reconciliation history, access control, and metadata live in a local SQLite database, or your own Postgres in production. Bring-your-own-key credentials are encrypted at rest.
Both. Engram is open source and runs locally by default, or routes remember, recall, improve and forget to a hosted Cognee Cloud tenant over REST, with automatic local fallback.
Give your AI a memory that lasts.
Sign in with GitHub or Google and start building a knowledge graph that catches contradictions the moment they appear and forgets only what has stopped mattering.