Engram

Engram gives an AI agent persistent memory across sessions by storing, searching, recalling and forgetting memories through semantic embeddings, entirely on the user's own hardware with no data leaving the machine. A single container bundles Qdrant, FastEmbed and an MCP server, and works with Claude Code, Cursor, Windsurf, VS Code, OpenClaw or any other MCP-speaking client.

Queries flow through three tiers in order: a hot-tier, in-memory frequency cache that answers in under a millisecond for memories accessed often, using an activation strength that grows with use and decays over time in the style of the ACT-R cognitive architecture; a multi-head locality-sensitive hashing index that takes the first 64 dimensions of a query embedding through four independent hash functions for constant-time candidate retrieval when the hot tier misses; and, beneath both, full vector search. The project positions itself against managed memory services that charge a recurring fee and route data through their own servers, and against simple local stores that do not scale past a few thousand memories or weight every memory equally.