run-llama/llama_index · error · ValueError

Token limit for full-text messages must be set and greater t

Error message

Token limit for full-text messages must be set and greater than 0.

What it means

ChatSummaryMemoryBuffer (which summarizes old messages instead of dropping them) uses the same pydantic validation pattern as ChatMemoryBuffer: token_limit must be present and >= 1, otherwise ValueError. The limit governs when full-text messages get summarized into a compressed summary.

Source

Thrown at llama-index-core/llama_index/core/memory/chat_summary_memory_buffer.py:72

    chat_store: SerializeAsAny[BaseChatStore] = Field(default_factory=SimpleChatStore)
    chat_store_key: str = Field(default=DEFAULT_CHAT_STORE_KEY)

    _token_count: int = PrivateAttr(default=0)

    @field_serializer("chat_store")
    def serialize_courses_in_order(self, chat_store: BaseChatStore) -> dict:
        res = chat_store.model_dump()
        res.update({"class_name": chat_store.class_name()})
        return res

    @model_validator(mode="before")
    @classmethod
    def validate_memory(cls, values: dict) -> dict:
        """Validate the memory."""
        # Validate token limits
        token_limit = values.get("token_limit", -1)
        if token_limit < 1:
            raise ValueError(
                "Token limit for full-text messages must be set and greater than 0."
            )

        # Validate tokenizer -- this avoids errors when loading from json/dict
        tokenizer_fn = values.get("tokenizer_fn")
        if tokenizer_fn is None:
            values["tokenizer_fn"] = get_tokenizer()

        return values

    @classmethod
    def from_defaults(
        cls,
        chat_history: Optional[List[ChatMessage]] = None,
        llm: Optional[LLM] = None,
        chat_store: Optional[BaseChatStore] = None,
        chat_store_key: str = DEFAULT_CHAT_STORE_KEY,
        token_limit: Optional[int] = None,

View on GitHub (pinned to afd0fef371)

Solutions

Use ChatSummaryMemoryBuffer.from_defaults(llm=llm, ...) which derives the limit from the LLM context window.
Pass token_limit explicitly as a positive integer.
Verify persisted dicts include token_limit before reloading.

Example fix

# before
memory = ChatSummaryMemoryBuffer(llm=llm)  # ValueError: no token_limit
# after
memory = ChatSummaryMemoryBuffer.from_defaults(llm=llm)
# or
memory = ChatSummaryMemoryBuffer(llm=llm, token_limit=3000)

Defensive patterns

Strategy: validation

Validate before calling

def is_valid_summary_memory_config(data: dict) -> bool:
    tl = data.get("token_limit", -1)
    return isinstance(tl, int) and tl >= 1 and data.get("llm") is not None

assert is_valid_summary_memory_config(config), "token_limit >= 1 and llm are required"

Type guard

def has_valid_token_limit(memory_dict: dict) -> bool:
    tl = memory_dict.get("token_limit", -1)
    return isinstance(tl, int) and tl >= 1

Try / catch

try:
    memory = ChatSummaryMemoryBuffer(llm=llm)
except ValueError as e:
    if "Token limit for full-text" in str(e):
        memory = ChatSummaryMemoryBuffer(llm=llm, token_limit=3000)
    else:
        raise

Prevention

Use from_defaults(llm=...) for ChatSummaryMemoryBuffer to auto-derive the limit.
Include token_limit whenever persisting summary memory config.
Remember this buffer also needs an llm for summarization.

When it happens

Trigger: Constructing ChatSummaryMemoryBuffer(...) directly without token_limit, with 0/negative, or deserializing a dict lacking the key — note it also requires an llm for summarization.

Common situations: Switching from ChatMemoryBuffer and assuming defaults exist on the raw constructor; loading persisted memory JSON that omitted token_limit; passing None explicitly.

Related errors

AI-assisted analysis of run-llama/llama_index@afd0fef371 (2026-08-15). Data as JSON: /api/errors/f176bdebf39d36cf. Report an issue: GitHub.