agent-memory

name: agent-memory description: Agent memory and context engineering capability pack. Gives AI agents the judgment rules for memory architecture (CoALA working/episodic/semantic/procedural layers), context compaction strategy selection, MemGPT/Letta virtual context management, Mem0 extract-reconcile pipelines, LangGraph state persistence and time-travel debugging, and Anthropic prompt-caching topology. Research-grounded rules from MemGPT/Letta, Mem0, LangGraph, the CoALA framework, and Anthropic caching docs. Use for any agent memory design, context-window optimization, checkpointing, or long-horizon statefulness task. keywords: ["记忆", "agent memory", "智能体记忆", "上下文工程", "context engineering", "compaction", "压缩", "checkpoint", "检查点", "MemGPT", "Letta", "Mem0", "prompt caching", "提示缓存", "CoALA", "时间旅行", "time travel", "LangGraph", "长期记忆", "context window", "上下文窗口"] type: reference-based

CONSUMES: User agent description + memory/context requirements + optional existing memory configs, checkpointer setup, or prompt structure PRODUCES: Applied memory-architecture judgment rules + compaction strategy selection + checkpointer/time-travel configs + caching breakpoint layout + memory-vs-vector-DB decision

Agent Memory & Context Engineering Capability Pack

Version: 0.1.0 Compatibility: Claude Code (Phase 1); Codex / Cursor / Gemini in Phase 3 License: Apache 2.0

What This Pack Does

AI agents build "memory" by appending every conversation turn to a vector store and calling it long-term memory. They expand the context window to fit raw history, paying quadratic attention cost. They summarize blindly at no fixed threshold, drifting and hallucinating. They store temporary events ("user bought coffee March 4") with the same weight as durable preferences ("user prefers black coffee"). They restart multi-step workflows from scratch on a crash because nothing was checkpointed. They place a timestamp at the top of the prompt and silently destroy every cache hit.

This pack embeds the judgment rules that memory and context engineers apply automatically — rules from MemGPT/Letta, Mem0, LangGraph, the CoALA framework, and Anthropic's caching documentation.

Pack = memory & context judgment. Your workflow system = process constraints. No overlap.

Cross-Cutting Rule: Memory System ≠ Vector Database

A vector database is a stateless similarity index; an agent memory system is a stateful architecture that governs the cognitive lifecycle of information — deciding what to retain, consolidate, modify, and discard over time. Appending every raw turn to an append-only vector store is NOT memory: it produces relevance drift and context dilution (near-duplicate entries flood the window), and vector similarity cannot do the multi-hop graph traversal real memory needs. A true memory layer MUST implement three cognitive processes on top of (or instead of) the vector store: Consolidation (dedupe/merge overlapping experiences), Scoring (importance weights + temporal decay so stale memories fade), and Temporal Tracking (index when facts change, e.g. Zep's Graphiti, so "used to code in Python" is distinguishable from "now codes in Rust").

Source: findings.md "Vector Storage vs. Stateful Memory Layers" [9, 31, 32, 33]

This rule applies to: memory architecture design, semantic-memory implementation, RAG-as-memory decisions, and any "store the conversation and search it later" proposal. It is surfaced here because burying it in one reference file causes agents to ship a vector store and call it done.

Quick Rule Index (all 34 rules)

One-screen map of every rule. Load the reference for full text + sources. Rule IDs are stable and citable in findings (e.g. [P0] Rule CE2).

Step 0: Context Detection

When the user mentions memory or context-engineering work, detect the context and load the right reference:

Step 1: Apply Rules

After loading the relevant reference file(s):

1. Read the reference completely — do not skim
2. Apply each rule as a judgment check against the user's memory design, config, or request
3. For each violated rule: state the violation clearly, then give the specific fix
4. Enforce the Memory System ≠ Vector Database cross-cutting rule on every "store and retrieve" proposal — demand consolidation + scoring + temporal tracking
5. Match the memory LAYER to the cognitive type — do not store a durable preference in a FIFO queue, or a system rule in episodic logs. Use the CoALA mapping in memory-architecture.md.

Output format per finding:

[P0] Rule MA2 (memory-architecture): Storing user preferences in the FIFO conversation queue — they evaporate when the session ends.
→ Move durable preferences to semantic memory (Mem0 user memory / Core Memory human sub-block), not the working-memory queue.

[P1] Rule PC3 (prompt-caching): Timestamp placed at top of the cached system prompt — invalidates the entire prefix every request.
→ Move dynamic variables AFTER the last cache_control breakpoint; keep Tools → System Prompt prefix byte-stable.

Step 2: Output

Produce a structured memory/context review:

## Memory & Context Review: [area reviewed]

### P0 — Blocking (must fix before shipping the agent)
- [finding + specific fix]

### P1 — Required (fix before production scale)
- [finding + specific fix]

### P2 — Advisory (improves cost/latency/fidelity)
- [finding + specific fix]

### Memory Layer Map
[table: each piece of state → CoALA layer → storage implementation]

### Tool Recommendation
[Letta / Mem0 / LangGraph checkpointer / Anthropic caching — based on user context]

Anti-Skip Table

Tool Quick Reference