Early-Signal Detector

method · first seen 4 Oct 2026

agent memory

Also: AI memory, agentic memory, AI agent memory, LLM agent memory, persistent memory, LLM memory systems, personal agent memory

Mechanisms by which LLM agents store, organize, and retrieve information from past interactions over long histories.

Agent memory covers how LLM agents write, store, organize and retrieve information across sessions and long task histories. The sources are mostly arXiv papers working on specific failure modes rather than the basic idea: memory-induced sycophancy, polluted or stale procedural memory in multi-agent systems, stale fact retrieval (a StaleBench benchmark), cost-aware memory for successive coding tasks, and program evolution to search memory designs evaluated on LoCoMo and LongMemEval. On GitHub, developers ship graph-database and bitemporal memory backends, MCP memory servers, and a ranked list comparing Mem0, Zep, Graphiti, Cognee and others. Weekly mentions have flattened and are now declining.

The concept itself is already mainstream, with established benchmarks and a crowded field of open-source products. The live work is narrower: making memory correct over time through staleness handling, verified outcomes and governance. That matters most for long-running coding and personal agents, which act on what they remember.

Analyst note (established), 17 Nov 2026.

Mentions per week

03419 May11 Aug10 NovPapersCodeDiscussion

PapersCodeDiscussion

Rank 39 on 14 Nov 2026 · 106 mentions in 28 days · 102 independent groups · papers → code → discussion · first seen 6 weeks ago · acceleration -0.21

Earliest sources

Latest sources

Mentioned together with

retrieval-augmented generationmodel context protocolmulti-agent LLM systemLoCoMo benchmarksycophancyBM25graph-based memoryLLM unlearningcontext engineeringspeaker identification