Agent Memory Optimizer

Agent Memory Optimizer

Overview

Treat memory as a time-phased system, not a single file or vector store. Analyze the project across four stages:

1. Before conversation: what is preloaded statically?
2. During conversation: what is surfaced dynamically?
3. After each turn: what is persisted durably?
4. Under context pressure: what preserves continuity during compaction?

Use this skill to audit the current implementation, design a target architecture, and implement the smallest phase-complete upgrades first.

Read references/claude-code-memory-patterns.md when designing or explaining the target architecture. Read references/audit-checklist.md when auditing an unfamiliar repo or producing a migration plan.

Workflow

1. Map The Current Memory System

Inventory the repo before proposing changes.

• Find prompt builders, user-context assembly, attachment injection, compaction, summaries, background hooks, and memory storage paths.
• Identify every place memory enters or leaves the model context.
• Distinguish rules/instructions from memory content.
• Distinguish cross-session durable memory from current-session notes.

Useful search terms:

rg -n "(memory|context|prompt|attachment|recall|retrieve|summary|compact|hook|session)" .

Produce a phase map in this shape:

• Startup preload
• Runtime dynamic recall
• Turn-end extraction
• Compaction/session continuity

2. Diagnose By Phase, Not By File

For each phase, answer:

• What triggers it?
• What data source does it use?
• What gets injected or written?
• What concurrency or dedupe guard exists?
• What token or byte budget exists?
• What are the obvious failure modes?

Do not jump straight into implementation. First find which phase is missing, overloaded, or conflated.

3. Spot High-Value Structural Gaps

Prioritize these architectural problems:

• Rules and memory content are mixed into one blob.
• All memory is preloaded at startup, with no runtime surfacing.
• Long-term memory and session continuity use the same store.
• Turn-end extraction has no cursor, dedupe, or overlap guard.
• Shared durable memory is written by many workers without a single-writer design.
• Compaction relies on ad hoc summarization instead of maintained session notes.
• One giant memory file stores everything instead of index plus topic files.
• File/path-local rules are missing, so all context must come from semantic recall.

4. Design The Target State

Use this upgrade order unless the repo proves a different bottleneck:

1. Split instruction memory from memory content.
2. Split durable memory from session memory.
3. Add structured storage: entrypoint index plus topic files.
4. Add runtime dynamic surfacing for the current query.
5. Add end-of-turn extraction with cursoring and coalescing.
6. Add session-memory-backed compaction.

Prefer the smallest viable upgrade that cleanly completes one missing phase.

5. Implement Incrementally

Use these implementation rules:

• Keep startup preload deterministic and cheap.
• Keep runtime recall narrow and zero-wait when possible.
• Keep turn-end extraction off the critical path.
• Keep session continuity local to the current conversation.
• Keep subagent memory isolated unless the design explicitly calls for shared scoped memory.
• Add gates, thresholds, and budgets before enabling automatic memory writes.
• Preserve existing behavior with fallbacks while introducing a new phase.

If you write code, state which phase each change belongs to.

6. Validate Behavior End To End

Validate with concrete scenarios rather than only unit reasoning.

Test at least these cases:

• Fresh session startup loads only the intended static memory.
• Runtime recall surfaces relevant memory for a query without duplicating prior injections.
• Turn-end extraction writes only newly learned durable information.
• Concurrent or rapid turns do not produce duplicate memory writes.
• Compaction preserves current work state and keeps recent unsummarized messages.
• Subagents do not accidentally mutate shared durable memory unless intended.

Output Shape

When using this skill, prefer producing these artifacts:

• A four-phase architecture map of the current system
• A current-vs-target gap list
• A prioritized upgrade plan with smallest safe first step
• Code changes grouped by phase
• A validation checklist tied to failure modes

Decision Rules

• If the repo only has startup memory, add runtime recall before building a more aggressive long-term memory writer.
• If the repo has durable memory but loses continuity after compaction, add session memory before improving semantic retrieval.
• If memory quality is poor because everything is stored together, fix storage structure before adding more retrieval sophistication.
• If automatic writing is risky, add manual or gated extraction first, then automate later.
• If the user asks for explanation rather than code, keep the output architectural and phase-based.