Collections

A collection is LangMem's default way of keeping memories: many separate memory documents, each with its own id, that grows as later conversations add to it.

Last updated: 30 Sep, 2026 · LangMem 0.0.30

One conversation gave one memory in Extracting memories. A customer comes back many times, so memories have to accumulate. A collection is that list; for each new conversation the manager decides whether to add a memory or change one it already has.

Syntax:

first = manager.invoke({"messages": conversation})
existing = [(m.id, m.content) for m in first]
later = manager.invoke({"messages": new_conversation, "existing": existing})

Passing what you already have

existing is a list of (id, content) pairs. The manager returns the whole collection after the conversation, old memories included, so you replace your list with the result.

existing = [(memory.id, memory.content) for memory in first]

A second conversation about a different order

later = manager.invoke({
    "messages": [{"role": "user", "content": "Order A-1002 is three days late."}],
    "existing": existing,
})

Two conversations with a single-call manager

from langchain.chat_models import init_chat_model
from langmem import create_memory_manager

model = init_chat_model("groq:openai/gpt-oss-120b", temperature=0)
manager = create_memory_manager(
    model,
    instructions="Extract what helps support this customer. Record everything in a single Memory call.",
)

first = manager.invoke({"messages": [
    {"role": "user", "content": "Hi, I'm Asha. Order A-1001 arrived broken. Please email me."},
]})
existing = [(memory.id, memory.content) for memory in first]

later = manager.invoke({
    "messages": [{"role": "user", "content": "Order A-1002 is three days late."}],
    "existing": existing,
})
for memory in later:
    status = "kept id" if memory.id in dict(existing) else "new id"
    print(f"{status:8} | {memory.content.content}")

kept id  | Customer Asha reported that order A-1001 arrived broken and requested to be contacted via email. Additionally, order A-1002 is three days late.

Why the collection did not grow

• kept id, and only one line. The second conversation did not add a memory. The model patched the existing one and appended "Additionally, order A-1002 is three days late."
• The instructions caused it. They asked for everything in a single Memory call, and the model kept everything in a single memory. It works like a single profile document, as in Profiles, and it only grows longer.
• Runs differ in how they merge. Another run wrote the merged text as a new memory next to the old one, which leaves a near-duplicate. Either way, A-1002 did not get a memory of its own.

Instructions that keep topics apart

The manager does what the instructions say, and "record everything in a single Memory call" reads as "keep one memory". Say how memories should be split, and a new order gets its own memory while the call stays single:

manager = create_memory_manager(
    model,
    instructions="Extract what helps support this customer. Record everything in a single Memory call. A new order or a new topic gets a new memory; do not merge it into an existing one.",
)

A new order in a new memory

from langchain.chat_models import init_chat_model
from langmem import create_memory_manager

model = init_chat_model("groq:openai/gpt-oss-120b", temperature=0)
manager = create_memory_manager(
    model,
    instructions="Extract what helps support this customer. Record everything in a single Memory call. A new order or a new topic gets a new memory; do not merge it into an existing one.",
)
first = manager.invoke({"messages": [
    {"role": "user", "content": "Hi, I'm Asha. Order A-1001 arrived broken. Please email me."},
]})
existing = [(memory.id, memory.content) for memory in first]

later = manager.invoke({
    "messages": [{"role": "user", "content": "Order A-1002 is three days late."}],
    "existing": existing,
})
for memory in later:
    status = "kept id" if memory.id in dict(existing) else "new id"
    print(f"{status:8} | {memory.content.content}")

new id   | Order A-1002 is three days late.
kept id  | Customer Asha reported that Order A-1001 arrived broken and requested to be contacted via email.

What the collection holds now

• new id for A-1002. The late order became its own memory, so the collection now holds two.
• kept id for A-1001: the memory from the first conversation is still there, under its own id.
• Only the instructions changed between this example and the one above; the model still made one tool call per reply.

Several rounds with max_steps

LangMem's reference documents a second way to get more memories out of one conversation: a max_steps key in the input. The manager then calls the model up to that many times; after each round it tells the model which memories were saved, and from the second round it adds a Done tool the model calls when it has finished. Each round is one more model call.

memories = manager.invoke({"messages": conversation, "max_steps": 3})

A conversation with nothing new

A collection only stays useful if nothing is saved twice. Send small talk with the existing memory and count:

from langchain.chat_models import init_chat_model
from langmem import create_memory_manager

model = init_chat_model("groq:openai/gpt-oss-120b", temperature=0)
manager = create_memory_manager(
    model,
    instructions="Extract what helps support this customer. Record everything in a single Memory call.",
)

first = manager.invoke({"messages": [
    {"role": "user", "content": "Hi, I'm Asha. Order A-1001 arrived broken. Please email me."},
]})
existing = [(memory.id, memory.content) for memory in first]

chat = manager.invoke({
    "messages": [{"role": "user", "content": "Thanks! Lovely weather in Pune today."}],
    "existing": existing,
})
print(len(existing), "before,", len(chat), "after")
for memory in chat:
    print(memory.content.content)

1 before, 2 after
Customer Asha reported that order A-1001 arrived broken and requested to be contacted via email.
Customer Asha reported that order A-1001 arrived broken and requested to be contacted via email.

1 before, 2 after. Nothing in the message was worth keeping, yet the manager returned a second memory, a word-for-word copy of the first with a new id. When existing memories are passed, LangMem's extractor binds the tools with tool_choice="any", so the model has to call one; here it re-created the memory it already had.

Collection vs profile

When a collection fits

• A customer's order problems, one memory per problem, kept even after the next one arrives.
• Notes a research assistant collects across many sessions.
• Anything where you cannot list the fields in advance.

Related

• Previous: Chat models
• Next: Updates and deletes
• Reference: create_memory_manager reference, with max_steps