Skip to contents

Agent memory policies

Source: R/memory.R
memory.Rd

An agent's memory decides which past messages enter the next context window. Three policies ship with the package; all are drop-in:

Usage

memory_buffer(keep = 40L)

memory_summary(threshold_chars = 12000L, keep_last = 10L, config = NULL)

memory_recall(embed_config, keep_recent = 6L, k = 4L)

Arguments

keep, keep_last, keep_recent, k, threshold_chars

Policy sizes; see above.

config

Optional LLMR::llm_config() used only for compaction summaries; NULL (default) summarizes with the agent's own model.

embed_config

An LLMR embedding config (e.g. llm_config("gemini", "gemini-embedding-001", embedding = TRUE)).

Value

A memory object to pass as agent(memory = ...).

Details

  • memory_buffer(keep): the last keep messages, verbatim. Simple, predictable, and right for most short-lived agents.

  • memory_summary(threshold_chars, keep_last, config): unbounded conversations. When stored text exceeds the threshold, the agent automatically condenses everything but the most recent messages into one summary note before its next request, so context stays small without forgetting decisions. By default the agent's own model writes the summary; pass config to bill compaction to a cheaper model instead.

  • memory_recall(embed_config, keep_recent, k): long-horizon recall. Older messages are embedded (via LLMR); at each turn the k most similar to the current input are injected alongside the recent tail.

Examples

m <- memory_buffer(keep = 10)
m$add("user", "hello")$add("assistant", "hi")
length(m$get())

if (FALSE) { # \dontrun{
cfg   <- LLMR::llm_config("groq", "openai/gpt-oss-20b")
cheap <- LLMR::llm_config("groq", "llama-3.1-8b-instant")

# an agent that summarizes its own past with the cheap model
scribe <- agent("Scribe", cfg,
                memory = memory_summary(threshold_chars = 8000,
                                        config = cheap))

# an agent that recalls relevant old exchanges by embedding similarity
emb <- LLMR::llm_config("gemini", "gemini-embedding-001", embedding = TRUE)
sage <- agent("Sage", cfg, memory = memory_recall(emb, k = 4))
} # }