Agent Memory

Memory is widely noted as one of the hardest problems in building practical LLM agents. LLM Agents — Research Overview frames the challenge directly: how does an agent retain and access information beyond what fits in a single prompt? [Research — LLM Agents.md]

There are two broad tiers:

• Short-term memory — the active context window. Fast and exact, but strictly limited in size. Everything the agent "knows" during a turn lives here.
• Long-term memory — information persisted outside the context and retrieved on demand. This is where most of the design complexity lives. [Research — LLM Agents.md]

A key claim in the research notes is that most systems advertised as having "agent memory" are, in practice, just retrieval + summarization — a Retrieval-Augmented Generation (RAG) pipeline feeding relevant chunks into the context window on each turn. [Research — LLM Agents.md]

This is functional but limited. The concern raised is that naive vector search misses temporal and structural relationships — for example, the order in which events happened, or hierarchical dependencies between pieces of information. A flat embedding lookup treats all stored facts as equally related and equally recent. [Research — LLM Agents.md]

*The typical RAG-based memory loop: long-term facts are retrieved, summarized, and injected into the context window alongside the current query.* [Research — LLM Agents.md]

There is an explicit conflict in the gathered sources on this point:

This is flagged in the original notes as an open disagreement worth investigating further. See Open Questions for the unresolved question this generates. [Research — LLM Agents.md]

Memory does not operate in isolation:

• In the ReAct Pattern, observations from tool calls are held in the context window (short-term memory) and accumulate across reasoning steps — making context management a live concern.
• Tool Use & Function Calling can itself be a memory mechanism: an agent can call a "memory write" or "memory read" function as an explicit tool action.
• Planner vs Reactive Agent Architectures is affected by memory design: explicit planners may store a full task decomposition in context, while reactive loops rely more heavily on retrieval to reconstruct state.

[Research — LLM Agents.md]

[Research — LLM Agents.md]