version: "1.0.0" name: devmemory-memory description: Search project memory before starting tasks, and store what you learn (architecture, decisions, gotchas) after making them. Use when starting a new codebase task, bug fix, or feature.

DevMemory: Shared Agent Memory

You have access to a shared project memory via the agent-memory MCP server. This memory persists across all sessions and is shared between every agent working on this project. Use it as a knowledgebase (look up past decisions) and coordination tool (leave context for future sessions).

1. Before Starting Any Task

Always search memory first. Before writing code, look up what's already known to prevent cold start:

search_long_term_memory(text="<describe what you're about to work on>", namespace="{{NAMESPACE}}")

Search for a specific topic:

search_long_term_memory(text="...", topics=["<topic>"], namespace="{{NAMESPACE}}")

Search for:

• Past decisions related to your task
• Known issues or gotchas in the area you're touching
• Established patterns and conventions
• Previous attempts that failed and why

Check for Hierarchical Summaries first before digging into commits:

# Query project-level summaries
search_long_term_memory(text="project summary architecture decisions", topics=["project-summary"], namespace="{{NAMESPACE}}", limit=3)
# Query architecture evolution summaries
search_long_term_memory(text="architecture evolution design patterns", topics=["architecture-summary"], namespace="{{NAMESPACE}}", limit=3)

2. After Making Significant Decisions

You have two ways to persist knowledge:

Structured knowledge: update .devmemory/knowledge/ files

For anything that future agents should know about, update the knowledge files directly and then sync:

When to update knowledge files:

• Architecture decision (add to architecture.md)
• Discovered a gotcha or workaround (add to gotchas.md)
• Established a new convention or pattern (add to conventions.md — create if needed)
• Added/changed a major dependency and why
• Fixed a non-obvious bug that could regress

Format for knowledge files:

---
topics: [architecture, decisions]
entities: [SQLite, Cloud API]
---
 
## Section Title
Content explaining what, why, and any relevant details.

Quick capture via MCP

After updating files, or for a single discovery, store via MCP:

create_long_term_memories(memories=[{
    "text": "<what was decided and why>",
    "memory_type": "semantic",
    "topics": ["<relevant>", "<topics>"],
    "entities": ["<technologies>", "<modules>"],
    "namespace": "{{NAMESPACE}}"
}])

3. Session Coordination

Use working memory to coordinate across active sessions:

When starting a large task — announce what you're working on:

set_working_memory(
    session_id="project-coordination",
    memories=[{
        "text": "Currently refactoring the search command to add LLM synthesis",
        "memory_type": "semantic",
        "topics": ["active-work"]
    }]
)