AutoMem's architecture is built on peer-reviewed memory research. This document tracks the four papers that most directly influenced the design, what each one found, and where in the codebase that finding shows up.
For the high-level pitch see the README. For the comparison against RAG and pure vector databases see COMPARISON.md. For benchmark methodology and history see benchmarks/EXPERIMENT_LOG.md.
Paper: HippoRAG 2: Smarter, Cheaper, and Better Memory for Stateful LLM Agents — Ohio State University, 2025.
Finding. A graph + vector hybrid retrieves associative memory roughly 7% better than pure vector RAG on multi-hop questions, and approaches human long-term memory performance on a battery of psychological recall tasks. The graph holds typed relationships; the vectors handle fuzzy semantic matching. Together they let an LLM agent surface the connections between memories, not just the closest match.
AutoMem implementation. AutoMem uses FalkorDB for the graph layer and Qdrant for the vector layer, with hybrid recall in automem/search/runtime_recall_helpers.py and graph traversal in automem/search/runtime_relations.py. Eleven authorable relationship types model the connections HippoRAG calls "associative" — LEADS_TO, EXEMPLIFIES, PREFERS_OVER, CONTRADICTS, and so on. Bridge discovery (multi-hop expansion via expand_relations) is the direct analogue of HippoRAG's "neighborhood" retrieval.
Paper: A-MEM: Agentic Memory for LLM Agents — 2025.
Finding. Memory benefits from dynamic organization: rather than fixed schemas, the system should evolve its own structure as memories accumulate, drawing on Niklas Luhmann's Zettelkasten method. Memories that are linked by recurring themes form emergent clusters; the cluster nodes themselves become a higher-order index that compresses the underlying graph.
AutoMem implementation. The enrichment pipeline (automem/enrichment/runtime_worker.py) extracts entities and themes from each new memory, then writes pattern nodes that other memories link into via EXEMPLIFIES edges. Consolidation cycles (automem/consolidation/runtime_scheduler.py) run weekly creative passes and monthly cluster passes that strengthen these emergent structures. The Pattern memory type is the first-class citizen for the resulting meta-knowledge.
Paper: MELODI: Exploring Memory Compression for Long Contexts — Google DeepMind, 2024.
Finding. A short-term + long-term memory hierarchy with selective gist compression preserves roughly 95% of recall quality at 8x lower memory cost than a flat retention strategy. The trick is deciding what to preserve in the gist — DeepMind shows that importance-weighted summarization beats uniform compression by a wide margin.
AutoMem implementation. Every memory above the soft length limit (MEMORY_CONTENT_SOFT_LIMIT) gets an automatic summary generated during enrichment. The decay cycle in automem/consolidation/runtime_scheduler.py runs an importance-weighted exponential, so high-importance memories with strong connections stay above the floor while low-importance memories with no inbound edges fade toward archival. The Forget cycle (off by default) is the long-term tail — archives memories below 0.2 relevance, deletes below 0.05.
Paper: A Human-Inspired Reading Agent with Gist Memory of Very Long Contexts — Google DeepMind, 2024.
Finding. Episodic memory and temporal organization extend an agent's effective context window roughly 20x, by letting the model "look up" the right page of history rather than carrying it all in working memory. Time-aware retrieval — knowing what came before, what came after, and what was contemporaneous — is at least as important as semantic similarity for long-horizon reasoning.
AutoMem implementation. Two relationship types carry temporal structure: OCCURRED_BEFORE (authorable) and PRECEDED_BY (auto-generated by the enrichment pipeline against recent memories). The recall scoring formula includes a temporal-alignment component and a recency boost, both tunable via SEARCH_WEIGHT_* env vars. time_query parameters on GET /recall accept natural-language ranges ("last 7 days", "last month") so an agent can scope retrieval to the right episode.
Add a new section when AutoMem incorporates a new technique from peer-reviewed work. Use the same four-part structure: paper link, finding (one paragraph), implementation (one paragraph with concrete file paths). Keep each section under 150 words so the doc stays scannable.
If you're considering a paper that influenced AutoMem's design but isn't yet implemented, open an issue tagged research rather than adding it here. This doc is for what we built, not what we might.