agent-memory

Agent Memory Protocol

3-tier persistent memory system for cross-session knowledge accumulation.

Memory Tiers

Tier	Path	Scope	Written By	Read By
Shared	.scaffolding/agent-memory/shared/KNOWLEDGE.md	Whole project	Any agent	All agents
Agent	.scaffolding/agent-memory/agents/{agent-name}/MEMORY.md	Per agent	Owning agent	Own agent + architect
Conversation	.scaffolding/conversations/{conversation_id}/agent-memory/context.md	Per conversation	Any agent in conversation	Agents in same conversation

Automatic Injection

Memory is auto-injected into agent context via recall_for_agent() in the task execution pipeline. When a task starts, the system reads:

1. .scaffolding/agent-memory/shared/KNOWLEDGE.md (always)
2. .scaffolding/agent-memory/agents/{agent-name}/MEMORY.md (when agent_name is known)
3. .scaffolding/conversations/{id}/agent-memory/context.md (when conversation_id is provided)

This means agents receive memory context automatically. Manual reading on first turn is optional but recommended for verifying latest data.

On First Turn

Before starting work, optionally read available memory for latest content (skip if files don't exist):

1. Read .scaffolding/agent-memory/shared/KNOWLEDGE.md
2. Read .scaffolding/agent-memory/agents/{your-agent-name}/MEMORY.md
3. If conversation_id is provided in task context: Read .scaffolding/conversations/{conversation_id}/agent-memory/context.md

Before Completing

Write significant findings to the appropriate tier:

Shared Knowledge (KNOWLEDGE.md)

Save here:

• Project architecture facts confirmed across multiple tasks
• Deployment gotchas and infrastructure quirks
• Cross-cutting patterns (e.g. how Redis is used, how tasks flow)
• Known bugs or limitations that affect multiple agents

Do NOT save:

• Agent-specific patterns (use agent memory)
• Task-specific context (use conversation memory)
• Anything already in CLAUDE.md or docs/

Agent Memory (MEMORY.md)

Save here:

• Patterns specific to your agent's domain (e.g. developer saves coding patterns)
• Lessons learned from mistakes in your domain
• File locations you frequently need
• Recurring debugging insights

Do NOT save:

• Generic project facts (use shared knowledge)
• One-time task details
• Speculative or unverified conclusions

Conversation Memory (context.md)

Save here:

• Decisions made during this conversation
• Findings from investigation (debugger -> developer handoff)
• Context needed by downstream agents in the same conversation chain
• Original intent and requirements clarifications

Do NOT save:

• Permanent knowledge (use shared or agent memory)
• Raw data or large code snippets

Read-Only Agents

Agents with disallowedTools: Write, Edit (architect, reviewer) cannot write to .scaffolding/agent-memory/ directly. These agents should report findings in their output, and writable agents in the same conversation chain can persist them.

Format Guidelines

• Max 200 lines per file (auto-injected memory has 200-line limit)
• Use markdown headers to organize by topic
• Include dates [YYYY-MM-DD] for time-sensitive entries
• Tag durable facts with confidence (high/medium/low) and last_verified [YYYY-MM-DD]; re-verify against live code/config before asserting a stale point-in-time fact, and down-grade or remove ones that no longer hold
• Remove outdated entries proactively
• Use concise bullet points, not prose

Hot/Cold Split

File-based memory (this skill) is the hot layer — auto-injected into every agent context under a 200-line budget, so keep it lean: stable, high-level facts and pointers only. Push detailed prose and rarely-needed, fuzzy-discoverable knowledge to the cold layer (vector store) via the semantic-memory-store skill — it only surfaces on similarity match and carries no per-turn token cost. Durable source-of-truth facts still get a file here; the cold copy is for natural-language recall.

Local-only carve-out: secrets and memory/MCP recovery procedures NEVER go to the cold vector store. Recovery info must stay readable when the store itself is down (a 401 means you cannot query the store to learn how to fix it), and secrets must not be embedded in a remote/shared backend. Keep these as file memory only.

File Creation

If memory files don't exist, create them with the appropriate header:

# Shared Knowledge
<!-- Cross-agent project knowledge. Max 200 lines. -->

# {Agent Name} Memory
<!-- Agent-specific patterns and lessons. Max 200 lines. -->

# Conversation {conversation_id} Context
<!-- Decisions and findings for this conversation chain. -->

Conversation Recall

Conversation memory is always available via file-based recall. When a task runs with a conversation_id, the system reads context.md and injects it into the agent's context. No database setup is required for conversation-tier memory.

Learning Loop

The /learn command closes the loop between a finished conversation and the memory tiers above. It distills one conversation into knowledge candidates (via the distill skill's Conversation-Scoped Distillation mode) and routes each candidate back into this memory system.

Candidate ingestion into the 3 tiers

Candidate kind	Target tier	File
Cross-cutting project fact	Shared	.scaffolding/agent-memory/shared/KNOWLEDGE.md
Domain-specific pattern or lesson	Agent	.scaffolding/agent-memory/agents/{name}/MEMORY.md
Decision relevant only to the active chain	Conversation	.scaffolding/conversations/{id}/agent-memory/context.md

Ingestion is propose-then-confirm: /learn prints the proposed memory diff and applies it only on explicit confirmation. Writes respect the 200-line KNOWLEDGE.md limit — overflow routes to the most relevant agent file per the distill overflow rule. Applied entries become auto-injected context on the next task.

Escalation to /create-skill

If a candidate is a repeatable procedure rather than a situational fact (per distill's Skill Promotion Criterion), it is not written to memory. Instead /learn proposes a /create-skill invocation with a pre-filled draft, escalating the knowledge into a first-class auto-invokable skill. Memory holds facts; skills hold reusable how-to procedures.