Everyone is building agent memory that remembers more. The five teams shipping production agents are building memory that forgets.
@Andre Lindenberg built MemoryBear with a pipeline modeled on human sleep consolidation: mark episodes for relevance, merge related ones into semantic knowledge, decay the rest. Additive-only memory degrades agent performance because stale memories poison the context window. Every append-only store eventually drowns the agent in outdated signal.
AnimaWorks implements this with ChromaDB for episodes and NetworkX graphs for spreading activation retrieval. The graph enables principled decay. When a memory node loses all high-weight connections, the system knows it can be forgotten. A vector in a flat store has no structural signal for obsolescence.
Five teams converge on what @Harrison Chase calls the "dreaming" pattern: batch-process execution traces during idle periods, extract patterns, synthesize into persistent knowledge. SimpleMem does it with a three-stage pipeline. OpenClaw evolves its SOUL.md file offline. @Jeel Patel's myworld CLI compiles scattered notes into 27 interconnected wiki pages. Each implementation differs. The principle is identical: learning happens when the agent is not working.
Chase frames this within a three-layer taxonomy. Agents learn at model weights, harness code, and runtime context. Catastrophic forgetting is a model-layer problem only. Harness changes are version-controlled. Context is append-only. But the context layer has its own failure mode: noise accumulation. Unbounded memory eventually hurts more than it helps.
XSkill resolves this by splitting knowledge into two streams. Experiences reduce tool errors by 45%. Skills improve task success by 20%. Both are needed. Experiences alone fix tool selection but break planning. Skills alone fix planning but miss tool nuances.
@Raphael Mansuy's SimpleMem adds a finding that rewrites retrieval assumptions: full text outperforms LLM-generated summaries by 47% F1. Summaries destroy signal and introduce hallucination at storage time. Preserve original text. Push abstraction to retrieval, not storage.
The connective layer across all of these is graph infrastructure. Graphs enable decay, spreading activation, contradiction detection, and entity resolution across sessions. Vector stores retrieve similar memories. Graphs retrieve structurally connected ones. The difference determines whether an agent compounds or drowns.
Resources:
- Agentic Graph RAG (O'Reilly): oreilly.com/library/view/agentic-graph-rag/9798341623163/
- MemoryBear: github.com/SuanmoSuanyangTechnology/MemoryBear
- AnimaWorks: github.com/xuiltul/animaworks
- XSkill: arxiv.org/abs/2603.12056
- SimpleMem: github.com/raphaelmansuy/autoResearch
- myworld CLI: github.com/jeelpatel2307/myworld
- Karpathy Wiki: karpathy.bearblog.dev