Memory & History

Manage conversation history and long-term memory in QitOS agents.

QitOS separates two distinct concerns:

• History — the conversation context sent to the LLM each step. It is ephemeral, message-scoped, and controlled by HistoryPolicy.
• Memory — a semantic record store that accumulates observations across steps and can be queried by strategy (window, summary, vector, or file).

Both are optional. If you pass neither, the Engine uses an internal 24-message window history and no memory adapter.

The mental model

When researchers say “context”, they often mean several different things at once. QitOS keeps them separate on purpose.

Concern	QitOS primitive	Typical use
What the model sees next step	History + HistoryPolicy	Prompt continuity
What the agent wants to remember across steps	Memory	Recall of findings, artifacts, and prior observations
What happens when context grows too large	TokenBudgetSummaryHistory / CompactHistory	Long-running control

This separation is one of QitOS’s most practical design choices for long-running agents. You can adjust message-window behavior without redesigning memory, and you can change memory strategy without changing the prompt protocol.

History

History stores HistoryMessage objects (role, content, step_id). The Engine appends a user message before each LLM call and an assistant message after. At call time, it uses HistoryPolicy to slice the window it passes to the model.

HistoryPolicy

HistoryPolicy lives in qitos.core.history and controls which messages the Engine selects:

from qitos.core.history import HistoryPolicy

policy = HistoryPolicy(
    roles=["user", "assistant"],  # message roles to include
    max_messages=24,               # hard cap on messages passed to the model
    step_window=None,              # if set, only include messages from the last N steps
    max_tokens=None,               # token budget ceiling (estimated)
)

Pass it to agent.run():

result = agent.run(
    task="...",
    max_steps=10,
    history_policy=policy,
    return_state=True,
)

HistoryPolicy.build_query() produces the retrieval query sent to the History adapter each step. The default Engine history implementation (_EngineWindowHistory) satisfies this query with a sliding window.

Tighten max_messages when your model has a small context window, or set step_window=4 to keep only the four most recent steps visible to the model.

Built-in history strategies

WindowHistory

WindowHistory is the simplest option. It keeps a fixed recency window and evicts older messages.

Choose it when:

• your runs are short
• you want easy-to-predict context behavior
• you do not want automatic summarization

TokenBudgetSummaryHistory

TokenBudgetSummaryHistory watches a token budget and summarizes older messages when the current window would exceed that budget.

Choose it when:

• the agent regularly runs long enough to hit context pressure
• you want a simple summary fallback
• you want token-budget behavior without the full multi-stage compaction system

CompactHistory

CompactHistory is the most advanced built-in context compactor.

It combines:

• warning thresholds before overflow
• grouping into compactable rounds
• micro-compaction for large old messages
• summary compaction for older interaction history
• runtime compaction metadata that can surface in traces

Choose it when:

• you are building long-running agents
• you care about inspectable context management
• you want context compaction to be part of the research artifact rather than hidden middleware

Context compaction in practice

CompactHistory is worth understanding because it captures one of QitOS’s most important design ideas: long-running behavior should be explicit and debuggable.

Its main pieces are:

• CompactConfig — thresholds and compaction policy
• MessageGrouper — how messages are grouped into rounds
• micro-compaction — trims very large old messages while keeping a preview
• summary compaction — replaces older rounds with a compact continuation summary

Typical setup:

from qitos.kit.history import CompactConfig, CompactHistory

history = CompactHistory(
    config=CompactConfig(
        max_tokens=16000,
        keep_last_rounds=2,
        keep_last_messages=8,
        warning_ratio=0.8,
        auto_compact=True,
    ),
    llm=llm,
)

This is usually the right choice when you are intentionally studying long-running tool use and want the context behavior to remain visible in traces.

Memory

Memory stores MemoryRecord objects (role, content, step_id, metadata). The Engine calls memory.append() after each observation and can retrieve records to inject into the prepare() context.

All four concrete adapters implement the same Memory ABC:

from qitos.core.memory import Memory, MemoryRecord

class Memory(ABC):
    def append(self, record: MemoryRecord) -> None: ...
    def retrieve(self, query=None, state=None, observation=None) -> list[MemoryRecord]: ...
    def summarize(self, max_items: int = 5) -> str: ...
    def evict(self) -> int: ...
    def reset(self, run_id=None) -> None: ...

Attaching memory to your agent

Pass a memory instance to the AgentModule constructor:

from qitos.kit.memory import WindowMemory

agent = MyAgent(
    llm=llm,
    memory=WindowMemory(window_size=20),
)

Or build it inline in __init__:

class MyAgent(AgentModule[MyState, dict, Action]):
    def __init__(self, llm):
        super().__init__(
            llm=llm,
            tool_registry=ToolRegistry(),
            memory=WindowMemory(window_size=20),
        )

---

WindowMemory

Keeps the most recent window_size records. Older records are dropped by evict().

from qitos.kit.memory import WindowMemory

memory = WindowMemory(window_size=20)

Retrieve with filters:

records = memory.retrieve(query={
    "roles": ["observation", "task"],  # filter by role
    "step_min": 3,                     # only records from step 3 onward
})

How eviction works:

evict() trims _records to the last window_size entries and returns how many were removed. The Engine does not call evict() automatically — call it yourself in reduce() if you need to actively prune:

def reduce(self, state, observation, decision):
    # ... update state ...
    if self.memory:
        self.memory.evict()
    return state

---

SummaryMemory

Retains the last keep_last records in memory. When evict() is called, older records are condensed into a one-line summary string stored in _summaries. Useful when you want the agent to have a compact reference to earlier work without consuming full context.

from qitos.kit.memory import SummaryMemory

memory = SummaryMemory(keep_last=10)

summarize() condenses the last max_items records into a pipe-separated string:

summary = memory.summarize(max_items=5)
# "observation:content of step N | assistant:..."

Access accumulated summaries via memory._summaries — these persist after eviction and can be injected into prepare():

def prepare(self, state):
    summaries = getattr(self.memory, "_summaries", [])
    context = "\n".join(summaries[-3:]) if summaries else ""
    return f"Task: {state.task}\nPast context:\n{context}"

---

VectorMemory

Retrieves records by semantic similarity using dot-product scoring. Each record is embedded on append() and ranked on retrieve().

from qitos.kit.memory import VectorMemory

# use the built-in character-bucket embedder (fast, low quality)
memory = VectorMemory(top_k=5)

# bring your own embedder (recommended for production)
def my_embedder(text: str) -> list[float]:
    return my_embedding_model.encode(text).tolist()

memory = VectorMemory(embedder=my_embedder, top_k=5)

Semantic retrieval:

records = memory.retrieve(query={
    "text": "what was the result of the last file read?",
    "top_k": 3,
})

When no text is provided, retrieve() falls back to returning the most recent top_k records.

The default embedder (_default_embedder) uses a character-bucket heuristic. It is fast and has no external dependencies, but its ranking quality is low. For meaningful semantic search, pass a real embedding function.

---

MarkdownFileMemory

Persists every record to a local Markdown file as an append-only log. The file survives restarts, making it useful for long-running or multi-session agents.

from qitos.kit.memory import MarkdownFileMemory

memory = MarkdownFileMemory(
    path="./workspace/memory.md",
    max_in_memory=200,  # how many records to keep in RAM
)

Each appended record is written as a Markdown section:

## Step 3 · observation
- time_utc: 2026-04-07T12:34:56+00:00

```text
{"action_results": [...]}

`retrieve()` supports `roles`, `step_min`, and `max_items` filters, same as `WindowMemory`. The file is never truncated — use `max_in_memory` to bound RAM usage while preserving the full on-disk log.

<Warning>
`reset()` clears in-memory records but does **not** delete or truncate the file. If you start a new run and reuse the same path, historical records from previous runs remain in the file.
</Warning>

---

## Convenience factories

`qitos.kit.memory` exports factory functions that mirror the constructors:

```python
from qitos.kit.memory import window_memory, summary_memory, vector_memory, markdown_file_memory

memory = window_memory(window_size=30)
memory = summary_memory(keep_last=8)
memory = vector_memory(top_k=4)
memory = markdown_file_memory(path="./runs/memory.md", max_in_memory=100)

---

Choosing a memory adapter

Adapter	In-memory	Persistent	Retrieval
WindowMemory	Yes	No	Recency + role filter
SummaryMemory	Yes	No	Recency; summarizes older records
VectorMemory	Yes	No	Semantic similarity
MarkdownFileMemory	Yes (bounded)	Yes	Recency + role filter

Start with WindowMemory for most agents. Switch to MarkdownFileMemory if you need persistence across restarts, or VectorMemory when you want the agent to recall semantically relevant past observations rather than just recent ones.

Recommended combinations

Goal	History	Memory
Small pattern demos	WindowHistory	none or WindowMemory
Medium coding agents	HistoryPolicy + default window or TokenBudgetSummaryHistory	WindowMemory or SummaryMemory
Long-running coding agents	CompactHistory	MarkdownFileMemory or SummaryMemory
Search/research agents	TokenBudgetSummaryHistory or CompactHistory	VectorMemory

Observe context behavior with qita

History and compaction are not meant to be invisible.

In long-running agents, inspect your runs with qita and look for:

• context-history warning events
• summary or compaction events
• whether later steps lose important constraints
• whether your chosen memory adapter is preserving the right artifacts

That is the research workflow QitOS is designed to support: tune context behavior, then inspect it as part of the run itself.