Continual Learning

Continual Learning

The ingestion pipeline for the PM-OS context graph. Reads chat transcripts, classifies what it finds, writes structured entries to memory/, and updates AGENTS.md with curated preferences and facts.

When to Use

• "Run continual learning" / "mine my chats" / "update memory"
• On a schedule (e.g. end of week) to keep the context graph current
• After a batch of important conversations (strategy sessions, PRD reviews, planning)

Inputs

• Transcript root: ~/.cursor/projects//agent-transcripts/
• Existing memory: AGENTS.md + memory/ subdirectories
• Incremental index: .cursor/hooks/state/continual-learning-index.json

Workflow

1. Load state

1. Read existing AGENTS.md.
2. Load incremental index (if present).
3. Discover transcript files. Process only:

• new files not in the index, or
• files whose mtime is newer than the indexed mtime.

2. Extract and classify

For each new/changed transcript, extract high-signal items and classify each into one of these categories:

| Category | What to look for | Write to | |----------|-----------------|----------| | Decision | User made or confirmed a product decision (PRD approval, scope change, launch/kill, prioritization shift, stakeholder alignment) | memory/decisions/ | | Exec update | User drafted or discussed an executive status update, SLT summary, or program update | memory/exec-updates/ | | Strategy shift | User discussed or changed strategic direction, pillars, positioning, or goals | memory/strategy-reviews/ | | Feedback insight | User analyzed customer feedback, VOC themes, or support trends | memory/feedback/ | | User preference | Recurring correction or stated broad rule about how the AI should behave | AGENTS.md → Learned User Preferences | | Workspace fact | Durable fact about the workspace, tools, file paths, or workflows | AGENTS.md → Learned Workspace Facts |

Skip: one-off task instructions, transient details (branch names, commit hashes, temp errors), secrets/tokens/credentials, and anything not actionable in future sessions.

3. Write to memory

For each extracted item (except preferences/facts), write a structured entry to the appropriate memory/ subdirectory:

Decisions → memory/decisions/YYYY-MM-DD_short-slug.md

# [Decision title]

**Date:** YYYY-MM-DD
**Source:** continual-learning (transcript extraction)
**Type:** [PRD approval | Scope change | Launch/Kill | Prioritization | Alignment | Other]

## Decision
[1-2 sentences: what was decided]

## Context
[What prompted this]

## Alternatives Considered
[If discussed in transcript, otherwise omit]

Exec updates → memory/exec-updates/YYYY-MM-DD.md

# Exec Update — YYYY-MM-DD

**Source:** continual-learning (transcript extraction)

## Summary
[Key points from the update]

Strategy shifts → memory/strategy-reviews/YYYY-MM-DD_topic.md

# Strategy: [topic]

**Date:** YYYY-MM-DD
**Source:** continual-learning (transcript extraction)

## What changed
[Description of the shift]

## Rationale
[Why, if discussed]

Feedback insights → memory/feedback/YYYY-MM-DD.md

# Feedback Insights — YYYY-MM-DD

**Source:** continual-learning (transcript extraction)

## Themes
[Themes identified]

Also write a log entry to memory/learning-log/YYYY-MM-DD.md:

# Continual Learning Run — YYYY-MM-DD

**Transcripts processed:** [count]
**Items extracted:** [count by category]

## Items written
- [category] → [path]: [1-line summary]

4. Update AGENTS.md

After writing to memory, update the Learned sections in AGENTS.md:

• Learned User Preferences — recurring corrections, stated broad rules
• Learned Workspace Facts — durable facts about workspace, tools, workflows

Rules:

• Update matching bullets in place (don't duplicate)
• Add only net-new bullets
• Deduplicate semantically similar bullets
• Max 12 bullets per section
• Plain bullet points only — no evidence tags, no metadata

5. Write back index

Update .cursor/hooks/state/continual-learning-index.json:

• Store latest mtimes for all processed files
• Remove entries for files that no longer exist on disk

Inclusion bar

Keep an item only if all are true:

• Actionable in future sessions
• Stable across sessions
• Repeated in multiple transcripts, or explicitly stated as a broad rule
• Non-sensitive

Exclusions

Never store:

• Secrets, tokens, credentials, private personal data
• One-off task instructions
• Transient details (branch names, commit hashes, temporary errors)

Incremental Index Format

{
  "version": 1,
  "transcripts": {
    "/abs/path/to/file.jsonl": {
      "mtimeMs": 1730000000000,
      "lastProcessedAt": "2026-02-18T12:00:00.000Z"
    }
  }
}

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: hardiktiwari
• Source: hardiktiwari/PM-operating-OS
• License: MIT

Install and usage instructions live in the source repository linked above.