| name | long-term-memory |
| description | Maintains a persistent project journal in MEMORY.md across sessions. Records state, architectural decisions, discoveries, task history, and open questions; reconciles with the project's planning documents; archives old entries automatically. Invoke at session start and after every completed task. Toggle with "memory on" / "memory off". Compatible with pair-mode. |
Long-Term Memory Management
The agent maintains a persistent long-term memory across sessions via a dedicated file: MEMORY.md. This file is the agent's journal — a durable record of what has been built, what has been learned, and what decisions were made (and why). It complements whatever product/planning documents the project uses by capturing the history and discoveries of the project.
Memory management is fully autonomous and ON by default: the agent reads, updates, reconciles, and archives MEMORY.md on its own initiative. The first-run bootstrap requires explicit user approval; everything after that is automatic until the user disables the mode.
🎚️ Memory Toggle
Memory is ON by default. The user can change state via:
memory off — disable writes (reads still happen for context).memory on — re-enable writes.memory status — report current state.
When OFF:
- The agent still reads
MEMORY.md at session start for grounding.
- The agent does NOT write, update, reconcile, or archive.
- If
MEMORY.md exists and memory is OFF, the agent reminds the user once per session: "📓 Memory is OFF; MEMORY.md exists but I won't update it. Say memory on to resume journaling."
State changes are announced explicitly:
📓 Memory: ON — I'll journal autonomously at task boundaries.
📓 Memory: OFF — I'll read MEMORY.md for context but won't write.
📂 Project Context Discovery
This skill is agnostic about which files describe the project's purpose and plan. At session start, the agent scans the project root (and one level deep) for files matching common conventions:
- Product/spec:
PRODUCT.md, README.md, SPEC.md, VISION.md, docs/product.md.
- Plan/roadmap:
PLAN.md, ROADMAP.md, TODO.md, BACKLOG.md, docs/plan.md.
- Architecture:
ARCHITECTURE.md, ADR/, docs/architecture.md.
Whichever exist become the source of truth for product intent and plan. If none exist, the agent asks the user to point at relevant docs or confirm there are none. Throughout this skill, "the plan document" and "the product document" refer to whatever the project actually uses.
📂 The MEMORY.md File
MEMORY.md lives at the project root. It is structured as an append-mostly log:
# Project Memory
## 📌 Current State Summary
A short (5–10 lines) snapshot of where the project stands right now. Rewritten in place after each completed task.
## 🏛️ Architectural Decisions
Durable decisions that shape the codebase. Append-only. Format: ADR-style, with a stable ID for cross-referencing.
- **ADR-001 — [One-line title]**
- **Context:** …
- **Decision:** …
- **Consequences:** …
## 🔍 Discoveries & Learnings
Non-obvious findings uncovered during implementation. Each entry has a stable ID for cross-reference.
- **DL-001 — YYYY-MM-DD — [short title]:** …
## 📜 Task Log
Chronological record of completed tasks. Append-only. Format:
`YYYY-MM-DD — [Task] — [Outcome] — [Files touched]`
## ❓ Open Questions & Follow-ups
Things noticed but not addressed. Moved to "Resolved" or deleted once handled.
Recommendation: commit MEMORY.md to version control — it is durable project history valuable to humans too. MEMORY.archive.md may optionally be .gitignored. The agent does not modify .gitignore autonomously.
🌱 First-Run Bootstrap
The first time the agent operates in a project, MEMORY.md will not exist. Before doing any other work:
Step 1: Detect
After attempting to read MEMORY.md, if missing:
- Announce:
🌱 No MEMORY.md found — bootstrapping long-term memory from [discovered files].
- Do NOT proceed with any task until bootstrap completes or the user declines.
Step 2: Gather
Read the discovered product/plan/architecture documents in full and extract:
- Project's purpose in 1–2 lines.
- Core domain concepts (for ubiquitous-language grounding).
- Already-accepted constraints.
- Current phase, next pending task, and any documented architectural decisions.
If documents are missing or too sparse, ask the user concise targeted questions before drafting.
Step 3: Draft
Compose an initial MEMORY.md using the section structure above. The header notes which files seeded it, e.g. Bootstrapped from README.md and ROADMAP.md on YYYY-MM-DD. Pre-existing decisions become ADR-001….
Step 4: Approve
Show the draft to the user and ask for explicit approval. Acceptable responses: looks good, approve, edits, or skip. On skip, set memory to OFF for the session and proceed.
🔄 Session-Start Protocol
At every session start (after bootstrap):
- Read
MEMORY.md in full.
- Skim Discoveries & Learnings so known gotchas surface during the upcoming work — cite
DL-NNN if a current task touches one.
- Recap to the user.
- Default: 2–3 lines covering current state, last task, next intended step.
- If the Current State Summary is unchanged from last session's end and recent: a one-line acknowledgement suffices (e.g. "📓 Memory loaded — picking up at task 3.2.").
- Stash an in-memory copy of the file's contents and its
mtime/hash for concurrent-edit detection (see below).
✍️ Update Protocol
Memory is updated at task boundaries. A task is "complete" when:
- The user confirms it, OR
- The working agreement's definition of done is met (e.g. tests green in TDD projects).
When in doubt, ask.
Concurrent-Edit Safety
Before every write, re-read MEMORY.md and compare against the stashed copy. If it changed:
- STOP. Do not write.
- Show the user a diff of their changes vs. the agent's intended update.
- Ask how to merge. Resume only after explicit direction.
What to record
- Task Log: always append one entry per completed task. Cap files-touched at 5; for more, summarize (e.g.
12 files across src/auth/).
- Current State Summary: rewrite in place to reflect new reality.
- Discoveries & Learnings: append entries that meet the heuristic below.
- Architectural Decisions: append a new ADR (with next sequential ID) only for durable, hard-to-reverse choices.
- Open Questions: append anything noticed but deferred; resolve or remove stale items proactively.
Discovery Heuristic
- ✅ Record anything that surprised you, contradicted docs, or would cost the next person >15 minutes to rediscover.
- ❌ Skip standard library behavior, things already in product/plan docs, or things obvious from reading the code.
- 🤔 Borderline: ask the user once early in the project; calibrate from their answer.
🔗 Reconciliation with the Plan
After updating memory, scan the plan document for drift and propose (do not auto-apply) reconciliations:
| Situation | Action |
|---|
| Plan task done but unmarked | Suggest marking it done. |
| Discovery invalidates a plan assumption | Flag the affected task; suggest a plan update. |
| New ADR conflicts with the plan | Surface the conflict; propose plan revision. |
| Renamed/moved task (same work, different name) | Ask the user to confirm equivalence; normalize naming in MEMORY.md going forward. |
| Open Question now answered | Move it to Resolved or delete it. |
The user owns all changes to the plan document.
🗄️ Archiving Policy
After each update, check thresholds. Archive if any holds:
- File exceeds 500 lines, OR
- Task Log exceeds 50 entries, OR
- Oldest Task Log entry is more than 90 days old.
When triggered:
- Create or append to
MEMORY.archive.md at the project root.
- Move the oldest half of Task Log and Discoveries & Learnings (preserving chronological order and IDs).
- Never archive Architectural Decisions, Current State Summary, or Open Questions.
- Leave a breadcrumb:
📦 Entries before YYYY-MM-DD archived to MEMORY.archive.md.
- Announce the archive operation to the user.
Surfacing Archived Context
When the agent encounters an unfamiliar domain term or decision not present in live MEMORY.md, it searches MEMORY.archive.md before asking the user. If found there, it cites the archived ID and (if still relevant) suggests promoting a brief note back into the live file.
📢 Announcing Memory Operations
Every memory-affecting action is announced compactly. Example:
✅ Task complete: UserProfileUpdated event implemented.
📝 Memory updated:
- Task Log: +1 entry (3 files touched)
- Discoveries: +1 entry (DL-014: event-sourcing library requires
@EventHandler on async handlers)
- Current State Summary: rewritten
🔗 Reconciliation: plan task 2.3 still marked pending — suggest marking it done.
🗄️ Archive check: 312 lines, 23 Task Log entries, oldest 41 days old. No archiving needed.
🤝 Interaction with Pair Programming Mode
When pair-mode is active, memory still works, with adaptations:
- Task Log entries credit the user:
YYYY-MM-DD — [Task] — ✅ Done (user-implemented, agent-guided) — [files].
- Discoveries surfaced by the user are tagged
(noted by user).
- The agent does NOT silently update memory mid-pair-session — updates happen at task boundaries, announced above, so the user owns the narrative.
- On
pair off: before pausing for user direction (per pair-mode protocol), flush a single memory update summarizing the pair session — one Task Log entry plus any discoveries — then wait for the user's next instruction.
📝 Writing Strategy
Prefer chunked, append-style writes over full-file rewrites where the host platform supports it. Always re-read before writing (see Concurrent-Edit Safety). When a full rewrite is necessary (e.g. updating Current State Summary), do so atomically and in a single operation.
