TaskManager Memory Skill

You manage the project-wide memory for this repository using SQLite.

Your goal is to:

1. Keep the memories table in .taskmanager/taskmanager.db valid and consistent.
2. Make it easy for other agents/skills/commands to discover relevant memories based on the current work.
3. Capture new long-lived knowledge (constraints, decisions, bugfixes, conventions) whenever it appears.
4. Track how often memories are used so the most important ones surface naturally.

Always work relative to the project root.

---

Database Location

• Database: .taskmanager/taskmanager.db
• Primary table: memories
• Full-text search: memories_fts (FTS5 virtual table)
• Task memory: state.task_memory (JSON column in state table)

Use sqlite3 via the Bash tool for all database operations.

---

Memory Schema

The memories table has these columns:

Column	Type	Description
id	TEXT PRIMARY KEY	Stable ID, e.g. "M-0001"
title	TEXT NOT NULL	Short summary (<= 140 chars)
kind	TEXT NOT NULL	One of: constraint, decision, bugfix, workaround, convention, architecture, process, integration, anti-pattern, other
why_important	TEXT NOT NULL	Concise explanation of why this memory matters
body	TEXT NOT NULL	Detailed description / rationale / examples
source_type	TEXT NOT NULL	One of: user, agent, command, hook, other
source_name	TEXT	Human/agent/command identifier
source_via	TEXT	Free-text, e.g. "cli", "tests/run-test-suite"
auto_updatable	INTEGER	0 for user-created (never auto-update), 1 for system-created
importance	INTEGER	1-5 (how critical), default 3
confidence	REAL	0-1 (how sure we are), default 0.8
status	TEXT	One of: active, deprecated, superseded, draft
superseded_by	TEXT	ID of newer memory (if superseded)
scope	TEXT (JSON)	Object with: project, files, tasks, commands, agents, domains. The tasks field links memories to specific task IDs for auto-loading during execution.
tags	TEXT (JSON)	Array of free-form tags, e.g. ["testing", "laravel"]
links	TEXT (JSON)	Array of links to docs/PRs/etc
use_count	INTEGER	Usage counter, default 0
last_used_at	TEXT	ISO timestamp of last use
last_conflict_at	TEXT	ISO timestamp of last detected conflict
conflict_resolutions	TEXT (JSON)	Array of conflict resolution history entries
created_at	TEXT	ISO timestamp
updated_at	TEXT	ISO timestamp

---

Note on Deferrals

Deferrals (tracked in the deferrals table) are separate from memories. Deferrals track work deferred from one task to another with source-target linkage and lifecycle management. They are managed by the run and update commands, not by this memory skill. Do not create memories to track deferred work; use the deferrals system instead.

---

Responsibilities

1. Initialize & Validate

When you start working:

1. Check if .taskmanager/taskmanager.db exists.
2. If not, the database needs initialization (use the taskmanager init process).
3. Verify the memories table exists:

   SELECT name FROM sqlite_master WHERE type='table' AND name='memories';
   

2. Query for Relevant Memories

Given a natural-language description of the current work (files, task IDs, domains):

1. Parse the description into:

   • Candidate domains (e.g. testing, performance, security, architecture).
   • Candidate files / directories.
   • Task IDs, if present.

2. Use SQL to find matching memories:

Full-text search (for keyword matching):

SELECT m.id, m.title, m.kind, m.why_important, m.importance, m.use_count
FROM memories m
JOIN memories_fts fts ON m.rowid = fts.rowid
WHERE m.status = 'active'
  AND memories_fts MATCH '<search_terms>'
ORDER BY rank, m.importance DESC, m.use_count DESC
LIMIT 10;

Scope-based search (for file matching):

SELECT id, title, kind, why_important, importance, use_count
FROM memories
WHERE status = 'active'
  AND (
    scope = '{}'
    OR json_extract(scope, '$.files') IS NULL
    OR EXISTS (
      SELECT 1 FROM json_each(json_extract(scope, '$.files')) f
      WHERE '<current_file>' LIKE f.value
    )
  )
ORDER BY importance DESC, use_count DESC
LIMIT 10;

Domain-based search:

SELECT id, title, kind, why_important, importance, use_count
FROM memories
WHERE status = 'active'
  AND EXISTS (
    SELECT 1 FROM json_each(json_extract(scope, '$.domains')) d
    WHERE d.value IN ('<domain1>', '<domain2>')
  )
ORDER BY importance DESC, use_count DESC;

Task-based search:

SELECT id, title, kind, why_important, importance, use_count
FROM memories
WHERE status = 'active'
  AND EXISTS (
    SELECT 1 FROM json_each(json_extract(scope, '$.tasks')) t
    WHERE t.value = '<task_id>'
  )
ORDER BY importance DESC;

3. Prefer:

   • Higher importance.
   • Higher use_count.
   • More recent last_used_at.

4. Return a compact summary (bullet list) with:

   • id, title, kind, why_important.
   • Any key constraints or decisions that MUST be respected.

You should never dump all memories into context unless explicitly asked; always select the smallest relevant subset.

3. Create a New Memory

When a user or another skill makes a decision that should persist for future work:

1. Check whether a similar memory already exists:

   SELECT id, title, kind FROM memories
   WHERE status = 'active'
     AND kind = '<kind>'
     AND (
       title LIKE '%<keyword>%'
       OR EXISTS (
         SELECT 1 FROM json_each(tags) t WHERE t.value = '<tag>'
       )
     );
   

2. If it is truly new, generate the next ID:

   SELECT 'M-' || printf('%04d', COALESCE(MAX(CAST(SUBSTR(id, 3) AS INTEGER)), 0) + 1)
   FROM memories;
   

3. Insert the new memory:

   INSERT INTO memories (
     id, title, kind, why_important, body,
     source_type, source_name, source_via, auto_updatable,
     importance, confidence, status,
     scope, tags, links,
     use_count, created_at, updated_at
   ) VALUES (
     '<id>', '<title>', '<kind>', '<why_important>', '<body>',
     '<source_type>', '<source_name>', '<source_via>', <0_or_1>,
     <importance>, <confidence>, 'active',
     '<scope_json>', '<tags_json>', '<links_json>',
     0, datetime('now'), datetime('now')
   );
   

When in doubt whether something deserves a memory, ask: "Will this decision/convention matter for future tasks?" If yes, create a memory.

Macro Decision Memory Example

When the plan command's macro architectural questions (Phase 3) capture a user decision:

INSERT INTO memories (
  id, title, kind, why_important, body,
  source_type, source_name, source_via, auto_updatable,
  importance, confidence, status,
  scope, tags
) VALUES (
  'M-0012',
  'Use Redis for queue driver',
  'architecture',
  'Affects all background job processing',
  'User chose Redis as the queue driver during macro analysis. Rationale: existing Redis infrastructure, supports priorities and delayed jobs.',
  'user', 'developer', 'taskmanager:plan:macro-questions', 0,
  4, 1.0, 'active',
  '{"domains": ["infrastructure", "queues"], "tasks": ["1.3", "2.1"]}',
  '["redis", "queue", "architecture"]'
);

Note the scope.tasks field linking this memory to relevant task IDs. During task execution, the run command auto-loads memories where the current task ID appears in scope.tasks.

4. Update or Supersede an Existing Memory

When an existing memory is refined or corrected:

1. If it's a small correction:

   UPDATE memories SET
     body = '<new_body>',
     tags = '<new_tags_json>',
     scope = '<new_scope_json>',
     updated_at = datetime('now')
   WHERE id = '<memory_id>';
   

2. If it's a substantial change or reversal:

   • Create a new memory entry with the updated decision.
   • Mark the old one as superseded:

   UPDATE memories SET
     status = 'superseded',
     superseded_by = '<new_id>',
     updated_at = datetime('now')
   WHERE id = '<old_id>';
   

Never silently rewrite history in a way that hides past decisions.

5. Track Usage

Whenever a memory directly influences planning or execution:

UPDATE memories SET
  use_count = use_count + 1,
  last_used_at = datetime('now'),
  updated_at = datetime('now')
WHERE id = '<memory_id>';

This allows future tools to treat highly-used, high-importance memories as more trustworthy.

---

How Other Skills/Commands Should Use This

When planning or executing non-trivial work (new features, refactors, risky changes):

1. Summarize the intent in a short description:
   • Files/directories involved.
   • Task IDs (if any).
   • Domains (e.g. testing, performance, architecture).
2. Use this skill to query for relevant memories.
3. Apply those memories as constraints and prior decisions when:
   • Creating task trees.
   • Designing architecture.
   • Writing or refactoring code.
   • Setting up tests, infra, or workflows.
4. If a new decision emerges that future work should follow:
   • Call this skill again to create or update a memory entry.

This way, the memories table becomes the single, durable "project brain" that all agents/commands/skills can rely on.

---

6. Task-Scoped Memory Management

Task-scoped memories are temporary memories that live only for the duration of a single task. They are stored in the state table under the task_memory JSON column.

6.1 Adding Task-Scoped Memory

When a user provides --task-memory "description" or -tm "description" to a command:

UPDATE state SET
  task_memory = json_insert(
    task_memory,
    '$[#]',
    json_object(
      'content', '<the description>',
      'addedAt', datetime('now'),
      'taskId', '<current task ID>',
      'source', 'user'
    )
  )
WHERE id = 1;

System-generated task memories use 'source', 'system'.

6.2 Retrieving Task-Scoped Memory

Before executing a task:

SELECT value FROM state, json_each(state.task_memory)
WHERE state.id = 1
  AND (
    json_extract(value, '$.taskId') = '<current_task_id>'
    OR json_extract(value, '$.taskId') = '*'
  );

Include these memories alongside global memories when applying constraints.

6.3 Promoting Task Memory to Global

At task completion (before marking "done"):

1. Check if any task memories exist for this task.

2. If task memories exist, use AskUserQuestion to ask:

   "The following task memories were used during this task. Should any be promoted to global (persistent) memory?"

   Options for each memory:

   • "Promote to global memory"
   • "Discard (task-specific only)"

3. For promoted memories:

   • Create a new global memory entry in the memories table.
   • Set source_type = 'user' if originally from user, or 'agent' if from system.

4. Clear the task memories for this task:

   UPDATE state SET
     task_memory = (
       SELECT json_group_array(value)
       FROM json_each(task_memory)
       WHERE json_extract(value, '$.taskId') != '<task_id>'
     )
   WHERE id = 1;
   

---

7. Conflict Detection (Opt-In)

Conflict detection is available via taskmanager:memory conflicts. It is NOT run automatically during task execution.

When invoked, it checks active memories for:

• File obsolescence: Referenced files in scope.files that no longer exist
• Stale memories: Active memories with use_count = 0 and created_at older than 30 days

Resolution

For each conflict found, offer resolution options via AskUserQuestion:

• "Keep as-is"
• "Update memory"
• "Deprecate memory"

Record all resolutions in the conflict_resolutions JSON column:

UPDATE memories SET
  conflict_resolutions = json_insert(
    conflict_resolutions,
    '$[#]',
    json_object(
      'timestamp', datetime('now'),
      'resolution', '<kept|modified|deprecated>',
      'reason', '<brief explanation>',
      'taskId', '<task ID>'
    )
  ),
  last_conflict_at = datetime('now'),
  updated_at = datetime('now')
WHERE id = '<memory_id>';

---

8. Conflict Resolution by Ownership

When resolving conflicts (whether via opt-in detection or during manual review):

User-Created Memories (source_type = 'user')

NEVER auto-update. ALWAYS ask the user via AskUserQuestion.

System-Created Memories (source_type != 'user')

• Refinements (small updates): Auto-update allowed.
• Reversals (substantial changes): Ask the user.

---

9. Memory Ownership Rules

9.1 Determining auto_updatable

When creating a memory, set auto_updatable based on source_type:

auto_updatable = (source_type != 'user') ? 1 : 0

• source_type = 'user' -> auto_updatable = 0
• source_type = 'agent' | 'command' | 'hook' | 'other' -> auto_updatable = 1

9.2 Update Rules by Ownership

Source Type	Small Update	Substantial Change
user	Ask user	Ask user
agent	Auto-update	Ask user
command	Auto-update	Ask user
hook	Auto-update	Ask user
other	Auto-update	Ask user

9.3 Never Delete Memories

Memories are never deleted. They are either:

• Kept as active
• Marked as deprecated (no longer relevant)
• Marked as superseded with a pointer to the new memory

This preserves decision history and audit trail.

---

10. Logging

All logging goes to a single file: .taskmanager/logs/activity.log.

<timestamp> [<level>] [memory] <message>

Levels: ERROR, DECISION. Logs are append-only.

Log these events:

• Memory created, updated, deprecated, superseded
• Memory applied to task
• Conflict detection and resolution
• Errors (database failures, invalid IDs, query errors)

Examples:

2025-12-11T10:00:00Z [DECISION] [memory] Created memory M-0005: "Always validate API inputs"
2025-12-11T10:00:01Z [DECISION] [memory] Applied memories to task 1.2: M-0001, M-0003, M-0005
2025-12-11T10:00:02Z [ERROR] [memory] Conflict: M-0001 references deleted file app/OldAuth.php
2025-12-11T10:05:00Z [DECISION] [memory] Deprecated M-0002: "No longer using old auth pattern"

---

SQL Quick Reference

Common Queries

Get all active memories:

SELECT * FROM memories WHERE status = 'active' ORDER BY importance DESC, use_count DESC;

Full-text search:

SELECT m.* FROM memories m
JOIN memories_fts fts ON m.rowid = fts.rowid
WHERE memories_fts MATCH '<search_term>'
  AND m.status = 'active'
ORDER BY rank, m.importance DESC;

Get memories for a specific file:

SELECT * FROM memories
WHERE status = 'active'
  AND (
    scope = '{}'
    OR json_extract(scope, '$.files') IS NULL
    OR EXISTS (
      SELECT 1 FROM json_each(json_extract(scope, '$.files')) f
      WHERE '<current_file>' LIKE f.value
    )
  )
ORDER BY importance DESC;

Record memory usage:

UPDATE memories
SET use_count = use_count + 1,
    last_used_at = datetime('now'),
    updated_at = datetime('now')
WHERE id = '<memory_id>';

Record conflict resolution:

UPDATE memories SET
  conflict_resolutions = json_insert(
    conflict_resolutions,
    '$[#]',
    json_object(
      'timestamp', datetime('now'),
      'resolution', '<resolution>',
      'reason', '<reason>',
      'taskId', '<taskId>'
    )
  ),
  last_conflict_at = datetime('now'),
  updated_at = datetime('now')
WHERE id = '<memory_id>';

Get next memory ID:

SELECT 'M-' || printf('%04d', COALESCE(MAX(CAST(SUBSTR(id, 3) AS INTEGER)), 0) + 1)
FROM memories;

Deprecate a memory:

UPDATE memories SET
  status = 'deprecated',
  updated_at = datetime('now')
WHERE id = '<memory_id>';

Supersede a memory:

UPDATE memories SET
  status = 'superseded',
  superseded_by = '<new_memory_id>',
  updated_at = datetime('now')
WHERE id = '<old_memory_id>';