Context Management

Context is the scarcest resource in every agent system. Four disciplines — compaction, selective retrieval, budget planning, memory safety — that keep long runs coherent.

What you'll learn

1. Context compaction
2. Selective retrieval
3. Context budget planning
4. Memory safety

Remember this

• Compaction preserves decisions, constraints, artifact pointers and open threads — triggered early, keeping recent turns verbatim
• Retrieve memory just-in-time with task anchors and thresholds; 'no memory' is a valid answer; conflicts resolved by provenance
• Budget every token: instrument layers, cap tool outputs, reserve reply headroom; never persist secrets or unscoped PII

Code: The context budget, enforced in one wrapper

BUDGET = {
    "system":  None,     # measured, fixed
    "memory":  2_000,    # tokens, retrieved top-k cap
    "recent":  8_000,    # verbatim turns
    "reserve": 4_000,    # reply + next tool-call headroom
}

async def build_context(state, task, user):
    ctx = [system_prompt(), tool_schemas()]              # fixed

    mems = memory.search(embed(task), k=4,               # selective
                         where={"user": user.id},        # access-checked
                         min_score=0.82)                  # abstain below
    ctx += format_memories(mems)[:BUDGET["memory"]]

    ctx += recent_turns(state, token_cap=BUDGET["recent"])
    if usage(ctx) > WINDOW - BUDGET["reserve"]:
        ctx = compact(ctx, keep_recent=3)                # ch.1 discipline
    return ctx

# Tool outputs are capped INSIDE each tool (AG-02): summarize-or-store,
# return pointer + summary. The wrapper never sees raw monsters.