Three Families of Memory Methods
knowledge-organization
structure what the model has thought or noted
retrieval-mechanism-oriented
an external library and a recall rule
architecture-driven
change the control flow that reads and writes
TiM
Think-in-Memory · Liu et al. 2023
raw turns
not kept
post-think
LLM reflects
after answering
thought
"user runs to get fit"
this is what
gets stored
bucket 1
bucket 2 ← thought
bucket 3
LSH recall
similar thoughts hash to
one bucket; a query reads
the bucket it lands in
memory ops on thoughts · insert · merge · forget
stores reasoning outcomes, not raw turns
recall by LSH over stored thoughts
memory ops · insert, forget, merge
focus ·
storage structure
A-Mem
Agentic Memory · Xu et al. 2025
earlier sessions
new session
note
kw · tags · ctx
note
kw · tags · ctx
note
kw · tags · ctx
updated ✎
links
new note
kw · tags · ctx
LLM writes it
memory evolution
a new note is linked to related notes and
may rewrite their context and tags
LLM writes a structured note per turn
links it into a knowledge network
old notes evolve when new ones arrive
focus ·
storage structure
MemoryBank
Zhong et al. 2024
vector DB · entries
daily conversation log
event summaries
user portrait · traits
each entry · embedding + time
memory strength
recalled → restored
time
Ebbinghaus forgetting curve
recall · query embedding vs entries; weak
entries fade out, recalled ones are refreshed
logs conversations, events, user traits
strength decays on an Ebbinghaus curve
refreshed on recall · portrait grows
focus ·
retrieval mechanism
MemGPT
Packer et al. 2023
LLM
main context · the fixed window
system prompt · recent messages · working set
page out
page in
LLM calls
external context · outside the window
recall storage (past messages) · archival storage
the model itself decides when to read and write
two contexts in an OS-like hierarchy
explicit read / write calls page data
past the fixed window
focus ·
control flow
four baselines
in the experiments · each family focuses on one dimension of memory management
06 / 25