YourMemory Adds Ebbinghaus Decay to Agent Memory

The GitHub repository for YourMemory (author: sachitrafa) publishes an agent-focused, persistent-memory project that applies the Ebbinghaus forgetting curve to decay stored memories and combines BM25, vector retrieval, and a graph layer for context-aware recall. The README reports test results on 1,534 QA pairs across 10 multi-session conversations with Recall@5 = 59% for YourMemory and shows a 28% baseline for Zep Cloud, presented as "2x better recall than Zep Cloud" in the README. The project documents a zero-infrastructure workflow using DuckDB for local storage and spaCy for NLP preprocessing, and provides MCP integrations for clients such as Claude, Cursor, and Cline (GitHub README).

The README attributes YourMemory's retrieval stack to a hybrid approach combining BM25, dense vectors, and a graph to surface related context rather than only direct matches. The project uses a decay model based on the Ebbinghaus forgetting curve to age memories and claims automatic pruning of outdated facts. The repository also includes a per-sample breakdown and an accompanying writeup for methodology (GitHub README). MCPMarket's product page restates these elements and adds performance claims including +16 percentage points recall improvement versus Mem0 on the LoCoMo benchmark and "100% stale memory precision" (MCPMarket listing).

Industry-pattern observations: combining sparse retrieval (BM25), dense embeddings, and a lightweight graph layer is a growing pattern for context-rich retrieval because each layer addresses different failure modes: lexical matches, semantic similarity, and relational context. Applying decay functions derived from human memory research such as the Ebbinghaus curve is an emerging heuristic for storage budgeting and freshness when persistent memory is local and uncurated. For practitioners: this hybrid-plus-decay design prioritizes recall of useful items while reducing retention of stale entities, which can lower hallucination risk in multi-session agents when implemented carefully.

Editorial analysis: open-source, local-first memory systems matter because they shift tradeoffs away from cloud-hosted state and back toward reproducible, auditable pipelines. YourMemory's reported Recall@5 = 59% on the cited benchmark and the comparative claims versus Zep Cloud and Mem0 are meaningful signals for teams evaluating agent memory stacks, but reported benchmark results should be verified against the full methodology and dataset, which the README says is available in the repository. Observers will note that zero-infrastructure installation (local DuckDB, spaCy) lowers adoption friction for developers who prefer on-device or self-hosted solutions.

Editorial analysis: watch for independent reproductions of the LoCoMo comparisons and for community feedback on multi-agent sharing semantics, stale-memory handling, and privacy properties of local DuckDB storage. Also monitor integration notes and edge-case behavior when the graph layer links many related memories; retrieval latency and index size are common operational pain points for hybrid stacks. If maintainers or independent reviewers publish benchmark code, practitioners should review the evaluation scripts and sampling to assess robustness.

Editorial analysis: for ML engineers building agents, YourMemory represents a low-friction prototype of persistent memory that pairs hybrid retrieval with biologically-inspired decay. The reported recall gains make it worth piloting in controlled settings, but teams should treat the benchmarked numbers as claims to be validated against their own conversational data and freshness requirements.