Genesys uses a causal graph, not a vector store, for long-term agent memory

StudentSweet3601 · reddit · 2026-07-21

Genesys is an open-source MCP memory server for agents, but it does not use a flat vector store. Instead, each memory is stored as a node in a causal graph with typed edges such as causedby, supports, contradicts, supersedes, derivedfrom, and temporalsequence.

Retention is scored multiplicatively by relevance, causal connectivity, and reactivation frequency, and a memory is only forgotten when it is low-relevance, causally orphaned, and unpinned at the same time.

The author says this makes it possible to inspect why something was remembered and to correct it explicitly: memoryamend creates a new node that supersedes the old one instead of overwriting history.

Current status:

pip install genesys-memory (AGPLv3)
Ships a local stdio MCP server
Works with Claude via a custom connector and is also in the ChatGPT directory
13 tools including store, recall, search, traverse, explain, amend, pin/unpin, and stats

The author invites people to break it and says the main bet is that causal-graph memory will beat vector-store memory once conversations get long.

Original post →

Genesys uses a causal graph, not a vector store, for long-term agent memory

More from coding & agent