Apam

APAM Memory Skill

Quick Commands

Command	When to use
/apam:apam-status	First step in any session that needs APAM - resolve the exact project_id
/apam:apam-fetch	Start of any session - load memory
/apam:apam-init	First time in a new project - bootstrap memory
/apam:apam-update	After building or designing something - save what was learned
/apam:apam-consolidate	Manually distill L2 episodes into L3 records

Use the focused commands above for day-to-day use. This document is the full policy reference.

You have access to persistent layered memory via the APAM MCP server. This memory persists across sessions - you never start blank. Follow this policy exactly and proactively. Do not wait to be asked to write memory. Write it as the knowledge is produced.

---

The Three Memory Layers

L1 - Fast Recall

Quick-access facts loaded into every session. Two categories:

Global (applies to all projects):

• User preferences ("prefers concise responses", "always use tabs")
• Communication style, tone, formatting preferences
• Workflow preferences ("always run tests before committing")

Project-specific (loaded only for this project):

• What the project is and what problem it solves (one sentence)
• Tech stack: languages, frameworks, key libraries
• Folder/module structure: what lives where
• Key entry points: main files, config files, server start command
• APIs and endpoints: names and one-line purpose each
• Databases and data models: what DB, what main schemas/collections
• Key external services or integrations
• Active constraints: "do not use X", "always go through Y"

L1 is an index, not a journal. Each atom is one fact, one sentence. Keep it scannable.

L2 - Episodes

A log of what happened during sessions. Written by Claude, not auto-generated. Each episode records a meaningful unit of work. You can write multiple episodes per conversation if multiple distinct things were accomplished.

Each episode should capture:

• What was done (2-4 sentence summary)
• Key decisions made
• Files significantly changed
• Problems fixed
• Patterns noticed
• If a plan was created: one-line summary plus path to the plan file
• If docs were created: one-line summary plus path to the doc file

L3 - Project Intelligence

Durable, structured knowledge about the project. Write to this directly and immediately using apam_update_intelligence; do not wait for episode consolidation. The periodic consolidation job will also update it, but direct writes are the primary path.

What belongs here:

Type	Example titles	Content
architecture	"System Overview", "Auth Design", "Data Pipeline"	How the system is structured, key design choices, why things are built the way they are
entity	"API Endpoints", "Database Schema", "Key Modules", "Integrations"	What exists: names, purposes, relationships - orientation-level, not full specs
procedural	"How to Run Tests", "Deployment Process", "Local Setup"	Step-by-step how-to knowledge needed repeatedly
pattern	"Error Handling", "Naming Conventions", "Test Strategy"	Recurring approaches, rules, conventions in the codebase

Also write Project Intelligence for:

• Future plans (architecture type, title "Future Plans"): things discussed but not yet built
• Known issues (entity type, title "Known Issues"): bugs or limitations noted
• Enhancement ideas (pattern type, title "Enhancement Ideas"): improvements discussed

---

Session Start Protocol

At the very start of every session, before doing anything else:

1. Run apam status in the current repository using the local shell.
2. Read the line that starts with Project: and copy the exact 16-character hex value.
3. Use that value as project_id in every subsequent APAM MCP tool call for this repository.
4. Call apam_recall with that project_id.
5. Read the returned context - it tells you the project, history, and what is already known.
6. If this is a new project with no memory:
   • Explore briefly: read package.json, README, and folder structure.
   • Or ask the user what the project is, what tech stack it uses, and what the goal is.
   • Then immediately pin what you learn to L1 and write initial Project Intelligence records.

Do not answer the user's request until you have loaded memory.

Critical rule: The project_id is the 16-character hex string on the Project: line of local apam status output. Never use a directory name, repo name, or any other string. Do not rely on MCP apam_status with no arguments for project discovery in a long-lived agent session, because the server cwd may not match the active repository. If you are unsure, run apam status again and read the value off the output.

---

During a Session: Write Memory As You Go

Write to L1 when:

New fact - something not covered by any existing atom:

• A user preference
• A new endpoint, module, service, or dependency
• A constraint ("never expose X", "this service is rate-limited")
• A confirmed decision

Existing atom is stale - this session changed something an atom describes:

• A tool was added or removed -> update the tools atom
• Folder structure changed -> update the structure atom
• Stack changed (new lib added, something swapped out) -> update the stack atom
• A constraint was lifted or changed -> update the constraint atom

Call apam_pin with the corrected full content. The upsert matches on type+content+scope, so write the complete updated fact - not just the delta. One fact per atom.

Write Project Intelligence immediately when:

• Architecture or design is discussed - even if nothing is built yet
• A new module, API, or service is designed or described
• You discover how something works that was not previously in memory
• Future plans or enhancements are discussed
• A document or plan is created (add a record pointing to it)
• A pattern or convention is established

Call apam_update_intelligence in the same response where the knowledge is produced. Do not defer to end of session.

Write an L2 episode when a meaningful chunk of work is complete:

• After implementing a feature
• After debugging and resolving an issue
• After creating a plan or design document
• After a significant design discussion that produced decisions

Multiple episodes per conversation is fine. Follow the work boundaries, not the conversation.

---

Populating L1 for a New Project

When you first encounter a project, pin these in order:

1. [decision/project] What this project is - one sentence
2. [decision/project] Tech stack (languages, frameworks, key libraries)
3. [decision/project] Key folder structure (where is src, tests, config)
4. [decision/project] Entry points (main file, CLI, server start command)
5. [decision/project] APIs/endpoints - names and one-line purpose each
6. [decision/project] Database/data layer - what DB, what main models
7. [decision/project] Key external services or integrations
8. [constraint/project] Any stated constraints or rules

Then write Project Intelligence records for architecture and entities with more detail.

---

Session End Protocol

Before your final response in any meaningful session:

1. Write any pending Project Intelligence not yet written mid-session
2. Write one or more L2 episodes
3. Pin any new L1 facts learned

---

Example: First Session on a New Project

User: "Let's build a REST API for task management using Node.js and PostgreSQL."

You:

1. Run apam status, then apam_recall using the exact project_id from the Project: line.
2. If memory is empty, pin to L1:
   • [decision/project] "Task management REST API - Node.js + PostgreSQL"
   • [decision/project] "Stack: Node.js, Express, PostgreSQL, TypeScript"
3. Write Project Intelligence immediately:

   apam_update_intelligence(type='architecture', title='System Overview',
     content='REST API for task management. Express server, PostgreSQL database,
     TypeScript. Users create, update, delete, and query tasks.')
   

4. Proceed with the work - as each design decision is made, write more records.
5. At the end, write an L2 episode.

---

Example: Updating Project Intelligence Mid-Session

User asks to add team endpoints. After designing them:

apam_update_intelligence(
  type='entity',
  title='API Endpoints',
  content='POST /tasks, GET /tasks, PATCH /tasks/:id, DELETE /tasks/:id - task CRUD
POST /teams, GET /teams/:id - team management
POST /teams/:id/members - add member
All routes require Bearer JWT.'
)

Do this in the same response where you design the endpoints.

---

Memory Hygiene

• No duplicates. If a fact is in recall output, do not re-pin - update the existing record instead.
• L1 atoms are granular. One fact per atom.
• Project Intelligence titles are keys. "API Endpoints" should update the same record rather than multiplying.
• If a Project Intelligence record is wrong, tell the user and suggest apam forget <card-id> before rewriting it with apam_update_intelligence.
• Trust user_confirmed L1 facts absolutely. Treat agent_inferred as strong defaults the user can override.
• Do not write episodes for trivial sessions - brief clarifying exchanges, no code written, no decisions made.