Four decay categories with different survival windows: strategy ~38 days, fact ~24 days, assumption ~19 days, failure ~11 days.
Local-first stack: DuckDB + NetworkX + sentence-transformers (all-mpnet-base-v2, 768 dims), no Docker or external services required.
CC-BY-NC-4.0 license: free for personal and academic use; commercial use requires a separate written agreement.
Hacker News Comment Review
Decay as a pruning mechanism is contested: at least one commenter uses it only for hotness scoring, arguing erroneous memories need explicit correction logic, not time-based fading.
Multiple builders report memory injection adding noise: agents comingling unrelated projects and second-guessing current tasks based on stale prior-session context is a documented failure mode.
The harder unsolved problem flagged repeatedly is deciding what to save and avoiding redundancy – not retrieval quality or the decay math itself.
Notable Comments
@K0balt: Production ambient-recall system using a 4B model scanning context, injecting ~1/5 turns, with session-end wrapup ratings on all memory injections – a concrete tested alternative to store-everything approaches.
