Module III — Context Engineering for Agents

Status: outline. Lecture body not authored.

Context stops being a prompt and becomes a system resource.

Topics

• context budgets and token allocation
• retrieval and relevance filtering
• compaction and structured summaries
• context contamination and leakage
• context inheritance across handoffs
• working context vs durable state
• memory admission gates and freshness policies

Memory, Context, State & Evidence

A rigorous harness enforces strict architectural separation between four distinct primitives:

MEMORY ≠ CONTEXT ≠ STATE ≠ EVIDENCE

• Context (Active Inference Context): The finite token window assembled specifically for an active inference pass. Subject to attention degradation and strict budget constraints.
• Memory (Persistent Informational Storage): Long-term, non-authoritative storage of past events, observations, or semantic facts across sessions. Memory *informs*, but cannot dictate execution.
• State (Operational Ground Truth): The authoritative, deterministic machine state (phase pointer, active step, variables, locks). State dictates *what the system is doing right now*.
• Evidence (Verifiable Observation): Verifiable records and artifacts used to substantiate claims about execution (tool outputs, process exit codes, hashes, test assertions).

The Memory Paradox

Core Principle:
More memory ≠ better reasoning.
Uncontrolled or incorrectly scoped memory can degrade inference.

Injecting retrieved memory into an LLM context is not cost-free. A retrieved memory item may be historically authentic and semantically similar, yet operationally toxic to the current reasoning step.

Failure Modes of Unbounded Memory

• Context Rot: As context expands with historical data, model attention across middle positions degrades (*lost in the middle*), obscuring immediate task constraints.
• Context Contamination: Precedents from previous iterations inject obsolete assumptions into fresh tasks.
• Excessive Context: High signal-to-noise ratio drops reasoning fidelity below single-shot zero-memory baselines.
• Stale Memory: Facts that were true at time $T_0$ become false at $T_1$ after code, environment, or requirement mutations.
• Cognitive Traps & Reasoning Fixation: The model anchors on past intermediate attempts, repeating flawed logic rather than exploring clean solutions.
• Belief Distortion: Semantic similarity in vector space measures embedding proximity, not operational authority, logical validity, or execution correctness.

Fundamental Axiom:
Retrieval relevance does not imply operational authority.

OLD VALID MEMORY (True at T0)
        ↓
    RETRIEVAL (High semantic similarity)
        ↓
 CURRENT CONTEXT (Injected without gate)
        ↓
    ANCHORING (Fixation on obsolete premise)
        ↓
WRONG CURRENT DECISION

Memory Admission

Before any retrieved memory candidate is admitted into active working context, the harness must apply an admission filter:

MEMORY CANDIDATE
      ↓
  scope check        (Does this memory apply to the active task domain?)
      ↓
freshness check      (Is this memory invalidated by newer events or clock?)
      ↓
authority check      (Who produced this memory? Is it an unverified claim or evidence?)
      ↓
contradiction check  (Does it conflict with active state or verified evidence?)
      ↓
WORKING CONTEXT      (Admitted into current inference window)

*(Control plane enforcement and automated policy gating are developed in Phase 7 · Module IX).*

Unique process — Context Fidelity Test

Evaluate context degradation and mutation across chained agent handoffs:

Agent A
   ↓  handoff
Agent B
   ↓
Agent C

The test measures five distinct vectors across handoff boundaries:

1. Information Survival: Percentage of essential task parameters retained. 2. Information Distortion: Semantic drift, hallucinated additions, or altered constraints. 3. Stale-State Propagation: Leakage of superseded or invalidated states downstream. 4. Contamination: Injection of irrelevant intermediate reasoning into clean agent contexts. 5. Authority Preservation: Maintenance of the boundary between verified facts and speculative inferences.

Retaining more tokens does not guarantee preserving operational fidelity. A harness that prunes aggressively often outperforms one that accumulates historical context indiscriminately.