When Facts Change: How Should an MCP Memory Server Update? Three Approaches Compared
izgorodin · reddit · 2026-09-12
The author poses a core test for memory systems: when a fact changes (project owner switches from Alice to Bob), can the agent replace the current fact while preserving history ("who owns it now" vs. "who owned it Tuesday")?
Three documentation-level paths are compared:
- Official MCP reference memory server: observations as explicit add/delete units — the caller decides what to remove; adding Bob never removes Alice.
- Basic Memory: Markdown files as the source of truth, editable via editnote or by humans directly; corrections are inspectable content edits.
- Graphiti: built-in fact validity windows with superseded-fact invalidation and date-filtered MCP fact search, pushing temporal semantics into the memory layer — though its MCP implementation is still experimental.
Suggested minimal test: after one handover, query the current owner in a fresh chat without old context attached, check the returned evidence, and test historical recall separately using the backend's time semantics. A fluent answer can hide an update that never happened. (Disclosure: the author co-founded Mnemoverse and excluded it from the comparison.)
More from coding & agent
- Developer says goodbye to VSCode after reassessing old habits in the AI era — jamestagg · 2026-09-12
- Hermes Agent uses new Rive CLI to rebuild homepage as a video game menu — Teknium · 2026-09-12
- Study: 69% of 31,000+ agent runs contained at least one reward-hacking episode — dair_ai · 2026-09-12
- Developer runs Codex entirely on a foldable phone, with app server and voice on-device — SIGKITTEN · 2026-09-12
- Cloud Agent Platforms Still Ask You to Pick a Repo While Codex Roams Unsupervised — willcb · 2026-09-12
- AI-generated firmware flashed straight to STM32 to run pumps and valve — debreuil · 2026-09-12