| name | agent-memory |
| description | Log agent sessions, consolidate project memory (dream), promote reviewed memory, and load context for new chats. Use when ending a session, improving cross-session recall, or when the user mentions agent memory, dream, log-session, or memory.md. |
Agent memory
Local consolidation for image-scoring-backend — external artifacts only (no model training).
When to use
- Session start: Read
.agent-memory/memory.md; prefer repo evidence on conflict.
- Session end / milestone: Log durable learnings with
/log-session.
- Periodic: Run
/dream-memory, review changelog, /promote-memory after human approval.
Memory candidates
Use text|category|confidence on CLI or YAML memory_candidates:
| Category | Use for |
|---|
stable_fact | Stack, architecture, env |
user_preference | Style, workflow, review prefs |
working_rule | How to run tests, what not to touch |
recurring_issue | Flakes, traps, env pain |
successful_pattern | Approaches that worked |
open_question | Unverified assumptions |
deprecated | Superseded guidance |
Confidence: low, medium, high.
Commands (repo root)
python scripts/agent-memory/log_session.py --summary "..." --outcome "..." --candidate "text|working_rule|high"
python scripts/agent-memory/import_transcripts.py --dry-run --cursor-projects "$env:USERPROFILE\.cursor\projects"
python scripts/agent-memory/dream.py
python scripts/agent-memory/promote_dream.py --dream .agent-memory/dreams/<timestamp>.md
python scripts/agent-memory/context.py
Transcript import (v2)
- Slash command:
/import-transcripts
- Config:
scripts/transcript_mining/workspace_map.json, repo_profiles.json
- Staging:
.agent/scratch/transcript-mining/<repo>/REVIEW.md (gitignored)
- Tier A:
--write-sessions → dream.py → human promote; Tier B/C: docs/LESSONS_LEARNED.md
- See docs/technical/AGENT_MEMORY.md § v2
Dream review checklist
- Open
dreams/*-changelog.md — scan Uncertain / needs review.
- Diff proposed
dreams/*.md vs memory.md.
- Redact anything sensitive; reject promotion if secrets might be present.
- Promote only when the proposal is accurate and concise.
Safety
- Scripts block common secret patterns on write.
- Do not log
secrets.json, .env, tokens, or personal library paths.
- Never edit
memory.md directly during implementation work.
Reference
