Zettelkasten Agent Memory Notes

Turn each interaction into an LLM-authored structured note — keywords, tags, a contextual description, and an embedding — instead of storing it under a developer-predefined schema.

Most agent memory systems (see memory management) require a developer to predefine memory storage structure up front — which fields exist, where each new fact is filed, when retrieval fires. That works until the agent encounters something the schema didn't anticipate: a novel solution the fixed categories can't file anywhere new, forcing it into whichever existing bucket is closest rather than the connections it actually has.

Zettelkasten-style agent memory notes avoid predefining that structure by having an LLM construct each memory note at write time. Each note is a structured tuple — original content, timestamp, plus three LLM-generated fields (keywords capturing key concepts, tags for categorization, and a contextual description giving richer semantic framing than the raw content alone) — with an embedding computed over all of it for similarity search. Generating the semantic fields with an LLM rather than a fixed schema lets the memory system extract implicit structure the raw interaction didn't state explicitly, and the multi-faceted representation (content + keywords + tags + description, not just a single summary) supports more nuanced retrieval later than a flat embedding of the raw text alone would.

This is a different bet from structured agent memory's developer-defined tables and queues, and from a plain memory stream's append-only natural-language log: the structure lives in per-note LLM-generated attributes decided at write time, not in a schema fixed at design time. The note itself is only half the picture — how a new note gets connected to existing ones is embedding-plus-LLM memory linking, and what happens to those existing notes once connected is memory evolution on retrieval.

Organization quality is bounded by the authoring LLM's own capability. Because the keywords, tags, and contextual description attached to each note are themselves LLM-generated rather than derived from a fixed procedure, a weaker model produces weaker organization — different LLMs can generate meaningfully different descriptions or draw different connections from the same underlying interaction. This is the cost side of trading a fixed schema for LLM-authored structure: the fixed schema had firm data at the cost of inflexibility, and letting an LLM decide structure trades that firmness for whatever quality ceiling the authoring model itself has.

This structure pays for itself in retrieval cost, not just organization quality. Because retrieval pulls a handful of compact, purpose-built notes (keywords, tags, and a short contextual description) instead of dumping large stretches of raw interaction history into context, one system reporting both architectures found the structured-note approach needed roughly an order of magnitude fewer tokens per memory operation than baselines that retrieve closer to full raw context — a large enough gap to change whether large-scale deployment is economically viable, not just a marginal efficiency win.

Sources

A-Mem (Xu et al.), §3.1, §4.3