saifyxpro Context Engineer
Design memory architectures and context window strategies for AI agents.
Workflow
Memory Design Workflow
- Identify what the agent needs to remember (facts, procedures, episodes)
- Classify memory into tiers using the three-tier model
- Select storage backend for each tier
- Define retrieval strategies and cache invalidation rules
- Set token budgets per context section
Context Audit Workflow
- Measure current context utilization (tokens per section)
- Identify redundant or stale content
- Apply the priority framework to rank sections
- Implement dynamic loading for low-priority knowledge
- Re-measure and compare
Three-Tier Memory Model
Read the relevant reference for implementation templates.
| Tier | What It Stores | Lifespan | Reference |
|---|---|---|---|
| Episodic | Specific interaction logs and outcomes | Session or cross-session | references/01-episodic-memory.md |
| Semantic | General knowledge and learned patterns | Persistent | references/02-semantic-memory.md |
| Procedural | Workflows, strategies, and refined processes | Persistent, versioned | references/03-procedural-memory.md |
Context Budgeting
Read the reference for budget allocation templates.
| Section | Priority | Budget | Reference |
|---|---|---|---|
| System Identity | Critical | Fixed (5-10%) | references/04-context-budgeting.md |
| Active Task Context | Critical | Dynamic (30-50%) | references/04-context-budgeting.md |
| Retrieved Knowledge | High | Dynamic (20-30%) | references/04-context-budgeting.md |
| Conversation History | Medium | Sliding window (10-20%) | references/04-context-budgeting.md |
| Cached Results | Low | Evictable (5-10%) | references/04-context-budgeting.md |
Dynamic Context Loading
Read the reference for loading pattern templates.
| Pattern | Description | Reference |
|---|---|---|
| Just-In-Time | Load knowledge only when task requires it | references/05-dynamic-loading.md |
| Prefetch | Predict and preload likely-needed context | references/05-dynamic-loading.md |
| Eviction | Remove low-relevance content when budget exceeded | references/05-dynamic-loading.md |
Context Priority Framework
When context window is full, evict in this order (lowest priority first):
- Cached tool outputs — regenerable on demand
- Old conversation turns — summarize instead of keeping verbatim
- Background reference material — reload from storage if needed
- Retrieved examples — keep only the most relevant
- NEVER evict — system identity, safety rules, active task state
Provider-Agnostic Context Layer
Separate context from model:
<context_layer>
<identity>[System prompt — model-independent]</identity>
<knowledge>[Retrieved facts — stored externally]</knowledge>
<state>[Task progress — persisted to DB/file]</state>
<history>[Conversation — sliding window]</history>
</context_layer>
<model_layer>
<provider>[OpenAI | Anthropic | Google | Local]</provider>
<model>[Specific model name]</model>
<token_limit>[Context window size]</token_limit>
</model_layer>
Switching providers requires ONLY changing the model layer. Context layer stays identical.
Anti-Patterns
- Context Stuffing — cramming everything into the prompt regardless of relevance
- Stateless Agent — no memory between sessions, relearns everything
- Stale Cache — cached information never expires, becomes incorrect
- Token Waste — verbose formatting consuming budget (XML when plain text suffices)
- Lost in the Middle — critical information buried in the center of long contexts
Validation Scripts
Validate context architecture with automated scoring (0-10):
python3 scripts/validate_context.py <config_file> [--strict]
Checks three-tier memory detection (episodic/semantic/procedural), token budgeting, eviction policies, and flags anti-patterns (unbounded injection, raw history dumping, no eviction).