YourMemory adds persistent, time-decayed agent memory
YourMemory is an open-source persistent memory layer for AI agents that combines cosine-similarity retrieval with an Ebbinghaus-inspired decay model. Memories are assigned an importance score at write time, decay by category over elapsed days, and get a recall boost when they are reused, with low-strength items automatically pruned. The repo positions the system as a practical Claude/MCP integration and reports a self-benchmarked gain over Mem0 on LoCoMo, plus strong stale-memory precision, while keeping the setup lightweight with SQLite by default and optional PostgreSQL/pgvector for larger deployments.
Hot take: this is a credible prototype for agent memory, but the real contribution is the pruning policy, not a fundamentally new retrieval stack.
- –The design is easy to reason about: similarity handles candidate retrieval, while strength handles time decay, importance, and repetition.
- –The category-based decay rates are sensible for agent memory, where facts, assumptions, failures, and strategies age differently.
- –The benchmark claim is interesting, but it is still self-reported on a narrow setup, so it needs independent validation before anyone should treat it as general proof.
- –The system’s practical appeal is in being shippable now: MCP support, SQLite default, and a simple install path lower adoption friction.
- –If this holds up in broader real-world usage, it could be a useful pattern for persistent memory layers in long-running agents.
DISCOVERED
12d ago
2026-03-31
PUBLISHED
12d ago
2026-03-31
RELEVANCE
AUTHOR
Sufficient_Sir_5414