memory infrastructure / edge-native
Engram gives AI agents accurate, fast, private long-term memory — with reranking that runs at the edge, not on a server you pay for separately.
Accuracy collapses at scale
Vector similarity doesn't understand which memory matters most. As memory grows, recall quality degrades silently.
Latency from cloud round-trips
Every recall requires a network call to a hosted vector database, then a second inference pass for reranking. For real-time agents, this is a hard blocker.
Privacy requires trust
Centralized memory architectures send sensitive user context to external servers for embedding, search, and reranking.
How Engram Works
Stage 1
Deduplicate Client-side WASM
catches duplicate content before it hits the network.
Stage 2
Embed & Store Content
is embedded and indexed via the /v3/documents API.
Stage 3
Retrieve Semantic
nearest-neighbor search returns a candidate set.
Stage 4
Rerank A cross-encoder
scores every query-candidate pair jointly at the edge. No server. No round-trip.
Stage 5
Consolidate The Louvain
graph clustering algorithm merges redundant memories and maintains graph accuracy over time.
Every stage runs at the edge. Nothing phones home to rerank.
The reranker that rescues correct answers.
Pure vector search returned the wrong answer first. The cross-encoder corrected it.
| Candidate | Vector Score | Reranker Score |
|---|---|---|
| Python tutorials (2nd place) | 0.814 | 0.298 |
| JavaScript fundamentals (3rd place) | 0.801 | 0.215 |
| Python basics for beginners (1st place) | 0.791 | 0.847 |
Works with every major AI framework.
MIT licensed. Self-host in one command.
Engram is fully open source. The entire retrieval pipeline — deduplication, reranking, graph consolidation — is auditable, forkable, and deployable on your own infrastructure.
MCP Integration
- —MCP server for Claude Desktop and Cursor
- —Claude Code plugin for persistent coding memory
- —Connector support for Notion, Google Drive, and OneDrive